host pat is set to cover all memory types by Xen, which is
necessary to support guest mtrr/pat, especially when device
is passthroughed with VT-d. However pat on bsp is not=20
recovered which could make assigned device defunct after S3
resume
Signed-off-by Kevin Tian <kevin.tian@intel.com>
on next use.
Apart from teh small performance gain of this, my primary motivation
is to avoid TLB flushes very early in boot, when the system is not yet
properly set up for cross-TLB shootdowns.
Signed-off-by: Keir Fraser <keir.fraser@citrix.com>
This makes TLB flushing on page allocation more conservative, but the
flush clock should still save us most of the time (page freeing and
alloc'ing tends to happen in batches, and not necesasrily close
together). We could add some optimisations to the flush filter if this
does turn out to be a significant overhead for some (useful)
workloads.
Signed-off-by: Keir Fraser <keir.fraser@citrix.com>
1. Do not touch deferred_ops in invalidate_shadow_ldt(), as we may
not always be in a context where deferred_ops is valid.
2. Protected the shadow LDT with a lock, now that mmu updates are not
protected by the per-domain lock.
Signed-off-by: Keir Fraser <keir.fraser@citrix.com>
Since page_list_move_tail(), page_list_splice_init(), and
page_list_is_eol() are only used by relinquish_memory(), and that
function can easily be changed to use more generic accessors, just
eliminate them altogether.
Signed-off-by: Jan Beulich <jbeulich@novell.com>
I've been asked too many times about a hint in xm man page on how to
'detach' from a guest console - aka Ctrl+]
Signed-off-by: Jim Fehlig <jfehlig@novell.com>
While in some places the guest requested flushes were already folded
into the deferred ops, this wasn't done consistently.
Also avoid using an uninitialized variable, even if doing so has no
correctness issue.
Signed-off-by: Jan Beulich <jbeulich@novell.com>
Avoid the call to check_freq() by default, since that function may
spin up to 1ms on certain systems without indicating any kind of severe
failure. This matches similar behavior in Linux.
Avoid doing a cross processor call in get_cur_val() if the current CPU
has its bit set in the mask passed in. Also use the local variable
'cpu' consistently, allowing to remove another local variable.
Signed-off-by: Jan Beulich <jbeulich@novell.com>
... in order to make them disappear from the final image. Of course
they could as well be removed altogether, but I assumed that whoever
added them had a reason to do so.
Signed-off-by: Jan Beulich <jbeulich@novell.com>
... by moving as much of the option processing into cpufreq code as is
possible, by folding the cpufreq_governor option into the cpufreq one
(the governor name, if any, must be specified as the first thing after
the separator following "cpufreq=xen"), and by allowing each
governor to have an option processing routine.
Signed-off-by: Jan Beulich <jbeulich@novell.com>
accessible. GRUB2 seems to like to stick them really high sometimes
(just below 4GB).
The 32-bit C code framework that this sets up can also be used for
other stuff in future:
* early cmdline parsing
* relocating multiboot modules so they too are guaranteed accessible
Its interaction with normal Xen start-of-day, and with the 16-bit
assembly trampoline, needs a bit of thought.
Signed-off-by: Keir Fraser <keir.fraser@citrix.com>
This patch fixes the following compilation error.
Since struct page_list_head is defined in mm.h, sched.h needs mm.h.
Other circular inclusions are sorted out.
> In file included from xen/include/asm-ia64/linux-xen/asm/smp.h:50,
> from xen/include/linux/smp.h:5,
> from xen/include/asm-ia64/linux/topology.h:33,
> from xen/include/asm-ia64/linux-xen/linux/gfp.h:6,
> from xen/include/asm/mm.h:11,
> from xen/include/xen/mm.h:90,
> from viosapic.c:35:
> xen/include/xen/sched.h:174: error: field page_list has incomplete
> type
> xen/include/xen/sched.h:175: error: field xenpage_list has
> incomplete type
Signed-off-by: Isaku Yamahata <yamahata@valinux.co.jp>
1. No reason for i386 not to use the same definitions as x64
2. No need for shadow_page_info names to hang around.
Signed-off-by: Keir Fraser <keir.fraser@citrix.com>
This also neatly eliminates (in certain cases) the need to reserve a
page between nodes, as the heap control structures will never be freed
and thus can serve as the needed separator.
It also adds an additonal fallback attempting to make use of the
wasted tail of a (group of) page(s) used for the _heap array by putting the
avail array there if it fits.
Signed-off-by: Jan Beulich <jbeulich@novell.com>
The previous check was flawed in that it checked
- for a single bit to be set instead of a range of bits to be clear
- a machine address against a mask supposed to be used on MFNs.
Signed-off-by: Jan Beulich <jbeulich@novell.com>
From all I can tell, this doesn't violate lock ordering as other
places call heap allocation functions from inside hd->mapping_lock.
Signed-off-by: Jan Beulich <jbeulich@novell.com>
By combining the overlay fields that are 8 bytes long (on x86-64) into
a union separate from the one used for the 4 byte wide fields, no
unnecessary padding will be inserted while at the same time avoiding
to use __attribute__((__packed__)) on any of the sub-structures (which
risks misaligning structure members without immediately noticing).
Signed-off-by: Jan Beulich <jbeulich@novell.com>
... combining the list entry members of both structures and removing
the artificial 'mbz' member (shadow code must keep the real underlying
member 'count_info' at zero for the lifetime of pages use as shadows).
This also fixes a latent issue with u.inuse._domain not getting
explicitly cleared before returning shadow pages to the domain heap -
it just so happened that this member turned out to be zero in all
(normal?) cases when a shadow page ends its life (but there were
neither build nor run-time assertions that this would actually be the
case). The bug got exposed by a subsequent patch changing the order of
fields in struct page_info.
Signed-off-by: Jan Beulich <jbeulich@novell.com>
Unless more than 16Tb are going to ever be supported in Xen, this will
allow reducing the linked list entries in struct page_info from 16 to
8 bytes.
This doesn't modify struct shadow_page_info, yet, so in order to meet
the constraints of that 'mirror' structure the list entry gets
artificially forced to be 16 bytes in size. That workaround will be
removed in a subsequent patch.
Signed-off-by: Jan Beulich <jbeulich@novell.com>
This is ia64 counter part of 19089:39517e863cc8.
This patch is preliminary for removing xenheap.
Signed-off-by: Isaku Yamahata <yamahata@valinux.co.jp>
Updates and checks on count_info and page owner can safely be
non-atomic.
This is ia64 counter part of 19088:055c589f4791.
Signed-off-by: Isaku Yamahata <yamahata@valinux.co.jp>
xenheap isn't pinned down now so that allocating p2m pages from
xenheap doesn't make sense anymore.
This is preparation for removing xenheap.
Signed-off-by: Isaku Yamahata <yamahata@valinux.co.jp>
This patch remove the following warning.
> memory.c:116: warning: implicit declaration of function 'guest_physmap_mark_populate_on_demand'
> memory.c:206: warning: implicit declaration of function 'p2m_pod_decrease_reservation'
Signed-off-by: Isaku Yamahata <yamahata@valinux.co.jp>
This patch adds a videoram parameter (4MB is the default) to pv guests
too: if the pv guest is actually a stubdomain then the videoram size
is added to the initial required available memory.
This way there is no risk that a stubdom fails to populate the
videoram because the RAM freed for the videoram by dom0 has already
been used to create the stubdom.
Signed-off-by: Stefano Stabellini <stefano.stabellini@eu.citrix.com>
Check the page refcount rather than the type count when deciding if a
page is still mapped.
This catches pages which are mapped by qemu; it also removes the need
for gnttab_is_granted().
Signed-off-by: George Dunlap <george.dunlap@eu.citrix.com>
Since changeset 17617 the block device protocols are checked to be
either file or phy. This has later been fixed to also include
tap. This is still not a correct check as there can be arbitrary
additional protocols. Before this check was added you could have a
block-xyz hotplug script and xyz would work as a protocol, now it is
refused. An example for this mechanism is the block-nbd script which
is included in the tree.
This patch changes the check to allow file, phy, tap and all protocol
names for which a block-$protocol hotplug script exists. This should
fix the last regressions introduced by the check.
Signed-off-by: Kevin Wolf <kwolf@suse.de>
Keir Fraser
Isaku Yamahata