in unstable (changeset aeaa3c83f6e5). Tested on a dual Opteron x86-64
machine with 4GB of memory and a tg3 modified to only DMA below 2GB.
- swiotlb_dma_supported() checked that the device DMA mask was equal
or bigger than 4GB, when in fact all that's needed is that the device
be able to DMA into the swiotlb aperture (1GB on my AMD x86-64
machine). Some device are actually only 31-bit DMA capable, so this
would've tripped them.
- On some platforms, PCI unmaps and syncs are nops, so there's no need to
keep track of the dma_addr they need after the initial mapping. The
DMA API supports this via the DECLARE_PCI_UNMAP_ADDR macros (see
Documentation/DMA-mapping.txt). Since the swiotlb code does make us of
the dma_addr for swiotlb_unmap_xxx and (more importantly)
swiotlb_dma_sync_xxx, we need to define them to something
meaningful.
Signed-Off-By: Muli Ben-Yehuda <mulix@mulix.org>
- page is moved next to store/console pages -- these are never put in the
free page pool by the domain
- xen_start_info is now a pointer to the original page instead of the copy
Signed-off-by: Christian Limpach <Christian.Limpach@cl.cam.ac.uk>
The driver exports xs_talkv to user-space, allowing programs running in
the domain to use the kernel's xenbus connection. Intended use is driver
domain device configuration where the driver domain is not the domain
running xenstored.
Signed-off-by: Christian Limpach <Christian.Limpach@cl.cam.ac.uk>
all register parameters are clobbered on return. Furthermore,
multicall arrays are clobbered on return from a multicall (only the
result field can be trusted!).
These restrictions are enforced by debug builds of Xen in the
following ways:
1. On completion of a multicall, the call array is overwritten
with garbage before writing in the result fields.
2. On completion of a hypercall, all parameter registers are
overwritten with garbage.
Signed-off-by: Keir Fraser <keir@xensource.com>
unpin", which is already implemented for x86_32. Since we now only pin
the root rather than any of the other levels, the overall performance
became better especially with workloads that require heavy memory
management operations.
On 8-way x86_64 xenlinux (dom0) the kernel build was improved by about
10% (using make -j32). Even a small setup like a UP HT system, I see
about 3% performance gain with kernel build (make -j4).
Lmbench also shows improvements in fork/exec/sh:
Processor, Processes - times in microseconds - smaller is better
--------------------------------------------------------------------
Host OS Mhz null null open slct sig sig fork exec sh =20
call I/O stat clos TCP inst hndl proc proc proc
--------- ------------- ---- ---- ---- ---- ---- ---- ---- ---- ----=20
Linux 2.6.12- 3786 1.13 1.36 3.93 6.04 10.5 1.43 4.33 536. 1446 3614
Linux 2.6.12- 3786 1.13 1.36 3.91 6.03 10.4 1.44 4.38 346. 1050 2831
Signed-off-by: Jun Nakajima <jun.nakajima@intel.com>
The root cause of bug 169 is, machine_to_phys_mapping, starting from
0xffff800000000000, is mapped using 2M pages. When the system has RAM
no more than 2G, only one 2M page is allocated and only one PDE entry is
created correspondingly, so calling mfn_to_pfn with mfn > 0x80000 will
overflow this 2M page and cause a unable handled kernel paging request.
The mfn > 0x80000 comes from PCI device I/O memory, here from AGP
display card when booting X server. Jun suggested to use something like
get_user() when accessing machine_to_phys_mapping.
Signed-off-by: Xin Li <xin.b.li@intel.com>
- keep count of free references
- let callbacks specify the number of references they need at least
- get rid of terminal reference
- simplify and more correct locking for callbacks
Signed-off-by: Christian Limpach <Christian.Limpach@cl.cam.ac.uk>
Fix start-of-day code and FOREIGN_FRAME() macro in light
of this. Also fix pfn_to_mfn() to mask off the foreign-frame
marker bit. The few places where we rely on that bit we read
out of the array directly, rather than using the pfn_to_mfn()
accessor macro.
This fixes, among other things, booting domU on PAE with swiotlb
disabled.
Signed-off-by: Keir Fraser <keir@xensource.com>
- add gnttab_free_grant_reference
- make gnttab_free_grant_references interface consistent with other gnttab
functions
- add support for callback when references are freed
Signed-off-by: Christian Limpach <Christian.Limpach@cl.cam.ac.uk>
this fixes users of virt_to_machine/machine_to_virt to
use virt_to_mfn/mfn_to_virt where that is more appropriate.
This should be a big step to improved PAE stability.
Signed-off-by: Keir Fraser <keir@xensource.com>
We don't want to parse paths we read out of the store to _construct_
other paths and tie down the store layout for no good reason.
Also require the frontend path to exist and be valid before creating
the device.
Signed-off-by: Christian Limpach <Christian.Limpach@cl.cam.ac.uk>
Backend directories are of form backend/<type>/<frontend-uuid>/<id>.
This extra level makes an asymmetry with front ends (device/type/id) and
are too long to make nice bus-ids under Linux, but the code isn't too
horrible, and an OS which only wants to be a guest doesn't need to implement
it. It allows backends to calculate the front-end path directly, which
is a nice benefit.
Also: subtype matching is tightened: they must match exactly. That way a future
variant which requires a specific driver will not match current drivers.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Christian Limpach <Christian.Limpach@cl.cam.ac.uk>
were allocated with xen_create_contiguous_region(). Such areas
do not require bounce buffers (swiotlb).
Signed-off-by: Keir Fraser <keir@xensource.com>
and multi-page DMA accesses. This is based on a preliminary patch
from Suresh Siddha at Intel.
Signed-off-by: Keir Fraser <keir@xensource.com>
Signed-off-by: Suresh Siddha <suresh.b.siddha@intel.com>
Linux has several instances of repeated code to do updates to a range
of PTEs. Mapping memory between domains in Xen also tends to need to
do this quite frequently, to ensure page tables have been constructed
and to look up PTE addresses when making mapping-related hypercalls.
This patch adds a generic PTE walk-and-fill operation that takes a
function pointer to call on leaf entries. direct_remap_area_pages()
is updated to use the new call, ass are abuses of
__direct_remap_area_pages.
This patch also introduces two new helper functions for working with
page tables when mapping memory between domains:
create_lookup_pte_addr() returns the machine address of a PTE,
allocating intermediate page tables as necessary. touch_pte_range()
ensures that page tables exist for a virtual address range.
Many of the existing linux page table operations (e.g. zap/remap/etc)
could be modified to use this interface, which would potentially
shorten up mm/memory.c a bit.
