usb: Add USB_RET_ADD_TO_QUEUE packet result code

This can be used by usb-device code which wishes to process an entire endpoint
queue at once, to do this the usb-device code returns USB_RET_ADD_TO_QUEUE
from its handle_data class method and defines a flush_ep_queue class method
to call when the hcd is done queuing up packets.

Signed-off-by: Hans de Goede <[email protected]>
Signed-off-by: Gerd Hoffmann <[email protected]>

usb: Rename __usb_packet_complete to usb_packet_complete_one

And make it available for use outside of core.c

Signed-off-by: Hans de Goede <[email protected]>
Signed-off-by: Gerd Hoffmann <[email protected]>

xhci: Add a xhci_ep_nuke_one_xfer helper function

Signed-off-by: Hans de Goede <[email protected]>
Signed-off-by: Gerd Hoffmann <[email protected]>

ehci: Retry to fill the queue while waiting for td completion

If the guest is using multiple transfers to try and keep the usb bus busy /
used at maximum efficiency, currently we would see / do the following:

1) submit transfer 1 to the device
2) submit transfer 2 to the device
3) report transfer 1 completion to guest
4) report transfer 2 completion to guest
5) submit transfer 1 to the device
6) report transfer 1 completion to guest
7) submit transfer 2 to the device
8) report transfer 2 completion to guest
etc.

So after the initial submission we would effectively only have 1 transfer
in flight, rather then 2. This is caused by us not checking the queue for
addition of new transfers by the guest (ie the resubmission of a recently
finished transfer), while waiting for a pending transfer to complete.
This patch does add a check for this, changing the sequence to:

1) submit transfer 1 to the device
2) submit transfer 2 to the device
3) report transfer 1 completion to guest
4) submit transfer 1 to the device
5) report transfer 2 completion to guest
6) submit transfer 2 to the device
etc.

Thus keeping 2 transfers in flight (most of the time, and always 1),
as intended by the guest.

Signed-off-by: Hans de Goede <[email protected]>
Signed-off-by: Gerd Hoffmann <[email protected]>

ehci: Detect going in circles when filling the queue

For ctrl endpoints Windows (atleast Win7) creates circular td lists, so far
these were not a problem because we would stop filling the queue if altnext
was set. Since further patches in this patchset remove the altnext check this
does become a problem and we need detection for going in circles.

Signed-off-by: Hans de Goede <[email protected]>
Signed-off-by: Gerd Hoffmann <[email protected]>

ehci: Speed up the timer of raising int from the async schedule

Often the guest will queue up new packets in response to a packet, in the
async schedule with its IOC flag set, completing. By speeding up the
frame-timer, we notice these new packets earlier. This increases the
speed (MB/s) of a Linux guest reading from a USB mass storage device by a
factor of 1.15 on top of the "Improve latency of interrupt delivery"
speed-ups, both with and without input pipelining enabled.

I've not tested the speed-up of this patch without the
"Improve latency of interrupt delivery" patch.

Signed-off-by: Hans de Goede <[email protected]>
Signed-off-by: Gerd Hoffmann <[email protected]>

ehci: Improve latency of interrupt delivery and async schedule scanning

While doing various performance tests of reading from USB mass storage devices
I noticed the following::
1) When an async handled packet completes, we don't immediately report an
   interrupt to the guest, instead we wait for the frame-timer to run and
   report it from there
2) If 1) has been fixed and an async handled packet takes a while to complete,
   then async_stepdown will become a high value, which means that there
   will be a large latency before any new packets queued by the guest in
   response to the interrupt get seen

1) was done deliberately as part of commit f0ad01f92:
http://www.kraxel.org/cgit/qemu/commit/?h=usb.57&id=f0ad01f92ca02eee7cadbfd225c5de753ebd5fce
Since setting the interrupt immediately on async packet completion was causing
issues with Linux guests, I believe this recently fixed Linux bug explains
why this is happening:
http://git.kernel.org/?p=linux/kernel/git/torvalds/linux.git;a=commitdiff;h=361aabf395e4a23cf554cf4ec0c0c6963b8beb01

Note that we can *not* count on this fix being present in all Linux guests!

I was hoping that the recently added support for Interrupt Threshold Control
would fix the issues with Linux guests, but adding a simple ehci_commit_irq()
call to ehci_async_bh() still caused problems with Linux guests.

The problem is, that when doing ehci_commit_irq() from ehci_async_bh(),
the "old" frindex value is used to calculate usbsts_frindex, and when
the frame-timer then runs possibly very shortly after ehci_async_bh(),
it increases the frame-timer, and thus any interrupts raised from that
frame-timer run, will also get reported to the guest immediately, rather
then being delayed to the next frame-timer run.

Luckily the solution for this is simple, this means that we need to
increase frindex before calling ehci_commit_irq() from ehci_async_bh(),
which in the end boils down to simple calling ehci_frame_timer() instead
of ehci_async_bh() from the bh.

This may seem like it causes a lot of extra work to be done, but this
is not true. Any work done from the frame-timer processing the periodic
schedule is work which then does not need to be done the next time the
frame timer runs, also the frame-timer will re-arm itself at (possibly)
a later time then it was armed for saving a vmexit at that time.

As an additional advantage moving to simply calling the frame-timer also
fixes 2) as the packet completion will set async_stepdown to 0, and the
re-arming of the timer with an async_stepdown of 0 ensures that any
newly queued up packets get seen in a reasonable amount of time.

This improves the speed (MB/s) of a Linux guest reading from a USB mass
storage device by a factor of 1.5 - 1.7 with input pipelining disabled,
and by a factor of 1.8 with input pipelining enabled.

Signed-off-by: Hans de Goede <[email protected]>
Signed-off-by: Gerd Hoffmann <[email protected]>

ehci: Set int flag on a short input packet

According to 4.15.1.2 an interrupt must be raised when a short packet
is received. If we don't do this it may take a significant time for
the guest to notice a short trasnfer has completed, since only the last td
will have its IOC flag set, and a short transfer may complete in an earlier
packet.

Signed-off-by: Hans de Goede <[email protected]>
Signed-off-by: Gerd Hoffmann <[email protected]>

ehci: Get rid of packet tbytes field

This field is used in some places to track the tbytes field of the token, but
in other places the field is used directly, use it directly everywhere for
consistency.

Signed-off-by: Hans de Goede <[email protected]>
Signed-off-by: Gerd Hoffmann <[email protected]>

uhci: Move checks to continue queuing to uhci_fill_queue()

Rather then having a special check to start queuing after the first packet,
and then another check for the other packets in uhci_fill_queue(), simply
check the previous packet beforehand in uhci_fill_queue()

Signed-off-by: Hans de Goede <[email protected]>
Signed-off-by: Gerd Hoffmann <[email protected]>

uhci: Properly unmap packets on cancel / invalid pid

Packets with an invalid pid, or which were cancelled have
usb_packet_map() called on them on init, but not usb_packet_unmap()
before being freed.

Signed-off-by: Hans de Goede <[email protected]>
Signed-off-by: Gerd Hoffmann <[email protected]>

Rename target_phys_addr_t to hwaddr

target_phys_addr_t is unwieldly, violates the C standard (_t suffixes are
reserved) and its purpose doesn't match the name (most target_phys_addr_t
addresses are not target specific).  Replace it with a finger-friendly,
standards conformant hwaddr.

Outstanding patchsets can be fixed up with the command

  git rebase -i --exec 'find -name "*.[ch]"
                        | xargs s/target_phys_addr_t/hwaddr/g' origin

Signed-off-by: Avi Kivity <[email protected]>
Signed-off-by: Anthony Liguori <[email protected]>

Merge remote-tracking branch 'qemu-kvm/memory/urgent' into staging

* qemu-kvm/memory/urgent:
  memory: abort if a memory region is destroyed during a transaction
  i440fx: avoid destroying memory regions within a transaction
  memory: Make eventfd adhere to device endianness

Merge remote-tracking branch 'awilliam/tags/vfio-pci-for-qemu-20121017.0' into staging

* awilliam/tags/vfio-pci-for-qemu-20121017.0:
  vfio-pci: Mark non-migratable
  vfio-pci: Fix debug build

usb-serial: only expose device in guest when the chardev is open

Signed-off-by: Gerd Hoffmann <[email protected]>
Signed-off-by: Anthony Liguori <[email protected]>

usb-serial: don't magically zap chardev on umplug

Signed-off-by: Gerd Hoffmann <[email protected]>
Signed-off-by: Anthony Liguori <[email protected]>

serial: add pci-serial documentation

Signed-off-by: Gerd Hoffmann <[email protected]>
Signed-off-by: Anthony Liguori <[email protected]>

serial: add 2x + 4x pci variant

Add multiport serial card implementation, with two variants, one
featuring two and one featuring four ports.

Signed-off-by: Gerd Hoffmann <[email protected]>
Signed-off-by: Anthony Liguori <[email protected]>

serial: add windows inf file for the pci card to docs

Signed-off-by: Gerd Hoffmann <[email protected]>
Signed-off-by: Anthony Liguori <[email protected]>

serial: add pci variant

So we get a hot-pluggable 16550 uart.

Signed-off-by: Gerd Hoffmann <[email protected]>
Signed-off-by: Anthony Liguori <[email protected]>

serial: split serial.c

Split serial.c into serial.c, serial.h and serial-isa.c.  While being at
creating a serial.h header file move the serial prototypes from pc.h to
the new serial.h.  The latter leads to s/pc.h/serial.h/ in tons of
boards which just want the serial bits from pc.h

Signed-off-by: Gerd Hoffmann <[email protected]>
Signed-off-by: Anthony Liguori <[email protected]>

Call MADV_HUGEPAGE for guest RAM allocations

This makes it possible for QEMU to use transparent huge pages (THP)
when transparent_hugepage/enabled=madvise. Otherwise THP is only
used when it's enabled system wide.

Signed-off-by: Luiz Capitulino <[email protected]>
Signed-off-by: Anthony Liguori <[email protected]>

Merge remote-tracking branch 'quintela/migration-next-20121017' into staging

* quintela/migration-next-20121017: (41 commits)
  cpus: create qemu_in_vcpu_thread()
  savevm: make qemu_file_put_notify() return errors
  savevm: un-export qemu_file_set_error()
  block-migration: handle errors with the return codes correctly
  block-migration:  Switch meaning of return value
  block-migration: make flush_blks() return errors
  buffered_file: buffered_put_buffer() don't need to set last_error
  savevm: Only qemu_fflush() can generate errors
  savevm: make qemu_fill_buffer() be consistent
  savevm: unexport qemu_ftell()
  savevm: unfold qemu_fclose_internal()
  savevm: make qemu_fflush() return an error code
  savevm: Remove qemu_fseek()
  virtio-net: use qemu_get_buffer() in a temp buffer
  savevm: unexport qemu_fflush
  migration: make migrate_fd_wait_for_unfreeze() return errors
  buffered_file: make buffered_flush return the error code
  buffered_file: callers of buffered_flush() already check for errors
  buffered_file: We can access directly to bandwidth_limit
  buffered_file: unfold migrate_fd_close
  ...

Merge remote-tracking branch 'qemu-kvm/memory/dma' into staging

* qemu-kvm/memory/dma: (23 commits)
  pci: honor PCI_COMMAND_MASTER
  pci: give each device its own address space
  memory: add address_space_destroy()
  dma: make dma access its own address space
  memory: per-AddressSpace dispatch
  s390: avoid reaching into memory core internals
  memory: use AddressSpace for MemoryListener filtering
  memory: move tcg flush into a tcg memory listener
  memory: move address_space_memory and address_space_io out of memory core
  memory: manage coalesced mmio via a MemoryListener
  xen: drop no-op MemoryListener callbacks
  kvm: drop no-op MemoryListener callbacks
  xen_pt: drop no-op MemoryListener callbacks
  vfio: drop no-op MemoryListener callbacks
  memory: drop no-op MemoryListener callbacks
  memory: provide defaults for MemoryListener operations
  memory: maintain a list of address spaces
  memory: export AddressSpace
  memory: prepare AddressSpace for exporting
  xen_pt: use separate MemoryListeners for memory and I/O
  ...

pci: honor PCI_COMMAND_MASTER

Currently we ignore PCI_COMMAND_MASTER completely: DMA succeeds even when
the bit is clear.

Honor PCI_COMMAND_MASTER by inserting a memory region into the device's
bus master address space, and tying its enable status to PCI_COMMAND_MASTER.

Tested using

  setpci -s 03 COMMAND=3

while a ping was running on a NIC in slot 3.  The kernel (Linux) detected
the stall and recovered after the command

  setpci -s 03 COMMAND=7

was issued.

Signed-off-by: Avi Kivity <[email protected]>

pci: give each device its own address space

Accesses from different devices can resolve differently
(depending on bridge settings, iommus, and PCI_COMMAND_MASTER), so
set up an address space for each device.

Currently iommus are expressed outside the memory API, so this doesn't
work if an iommu is present.

Signed-off-by: Avi Kivity <[email protected]>

memory: add address_space_destroy()

Since address spaces can be created dynamically by device hotplug, they
can also be destroyed dynamically.

Signed-off-by: Avi Kivity <[email protected]>

dma: make dma access its own address space

Instead of accessing the cpu address space, use an address space
configured by the caller.

Eventually all dma functionality will be folded into AddressSpace,
but we have to start from something.

Reviewed-by: Anthony Liguori <[email protected]>
Signed-off-by: Avi Kivity <[email protected]>

memory: per-AddressSpace dispatch

Currently we use a global radix tree to dispatch memory access.  This only
works with a single address space; to support multiple address spaces we
make the radix tree a member of AddressSpace (via an intermediate structure
AddressSpaceDispatch to avoid exposing too many internals).

A side effect is that address_space_io also gains a dispatch table.  When
we remove all the pre-memory-API I/O registrations, we can use that for
dispatching I/O and get rid of the original I/O dispatch.

Signed-off-by: Avi Kivity <[email protected]>

s390: avoid reaching into memory core internals

use cpu_physical_memory_is_io() instead.

Signed-off-by: Avi Kivity <[email protected]>

memory: use AddressSpace for MemoryListener filtering

Using the AddressSpace type reduces confusion, as you can't accidentally
supply the MemoryRegion you're interested in.

Reviewed-by: Anthony Liguori <[email protected]>
Signed-off-by: Avi Kivity <[email protected]>

memory: move tcg flush into a tcg memory listener

We plan to make the core listener listen to all address spaces; this
will cause many more flushes than necessary.  Prepare for that by
moving the flush into a tcg-specific listener.

Later we can avoid registering the listener if tcg is disabled.

Signed-off-by: Avi Kivity <[email protected]>

memory: move address_space_memory and address_space_io out of memory core

With this change, memory.c no longer knows anything about special address
spaces, so it is prepared for AddressSpace based DMA.

Reviewed-by: Anthony Liguori <[email protected]>
Signed-off-by: Avi Kivity <[email protected]>

memory: manage coalesced mmio via a MemoryListener

Instead of calling a global function on coalesced mmio changes, which
routes the call to kvm if enabled, add coalesced mmio hooks to
MemoryListener and make kvm use that instead.

The motivation is support for multiple address spaces (which means we
we need to filter the call on the right address space) but the result
is cleaner as well.

Signed-off-by: Avi Kivity <[email protected]>

fix CONFIG_QEMU_HELPERDIR generation again

commit 38f419f35225 fixed a breakage with CONFIG_QEMU_HELPERDIR
which has been introduced by 8bf188aa18ef7a8.  But while techinically
that fix has been correct, all other similar variables are handled
differently.  Make it consistent, and let scripts/create_config
expand and capitalize the variable properly like for all other
qemu_*dir variables.

Signed-off-by: Michael Tokarev <[email protected]>

target-sparc: fix FMOVr instruction

Like the MOVr instruction, the FMOVr instruction has the condition
encoded between bits 10 and 12.

Cc: Blue Swirl <[email protected]>
Signed-off-by: Aurelien Jarno <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

hw/pl031: Use LOG_GUEST_ERROR

Use LOG_GUEST_ERROR rather than hw_error or direct fprintf.

Signed-off-by: Peter Maydell <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

hw/pl022: Use LOG_UNIMP and LOG_GUEST_ERROR

Use LOG_UNIMP and LOG_GUEST_ERROR where appropriate rather
than hw_error().

Signed-off-by: Peter Maydell <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

hw/pl011: Use LOG_UNIMP and LOG_GUEST_ERROR

Use the new LOG_UNIMP and LOG_GUEST_ERROR logging types rather
than hw_error().

Signed-off-by: Peter Maydell <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

hw/pl190: Use LOG_GUEST_ERROR

If the guest attempts an offset to a nonexistent register, just
log this via LOG_GUEST_ERROR rather than killing QEMU with a hw_error.

Signed-off-by: Peter Maydell <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

hw/pl041: Use LOG_UNIMP

Use the new LOG_UNIMP tracing to report unimplemented
features.

Signed-off-by: Peter Maydell <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

hw/pl181: Use LOG_UNIMP and LOG_GUEST_ERROR

Rather than a mix of direct printing to stderr and aborting
via hw_error(), use LOG_UNIMP and LOG_GUEST_ERROR.

Signed-off-by: Peter Maydell <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

hw/hw.h: Add include of qemu-log.h

Add an include of qemu-log.h to hw.h, so that device model
code has access to these logging functions without the need
to directly include qemu-log.h.

Signed-off-by: Peter Maydell <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

qemu-log: Add new log category for guest bugs

Add a new category for device models to log guest behaviour
which is likely to be a guest bug of some kind (accessing
nonexistent registers, reading 32 bit wide registers with
a byte access, etc). Making this its own log category allows
those who care (mostly guest OS authors) to see the complaints
without bothering most users.

Signed-off-by: Peter Maydell <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

target-sparc: Remove cpu_tmp0 as a global

Subroutines do their own local temporary management.
Within disas_sparc_insn we limit the existance of the variable
to OP=2 insns, and delay initialization as late as is reasonable
for the specific XOP.

Signed-off-by: Richard Henderson <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

target-sparc: Make cpu_dst local to OP=2 insns

And initialize it such that it (may) write directly to rd.

Signed-off-by: Richard Henderson <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

target-sparc: Only use cpu_dst for eventual writes to a gpr

Use cpu_tmp0 for other stuff, like Write Priv Register.

Signed-off-by: Richard Henderson <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

target-sparc: Remove last uses of cpu_tmp64

Signed-off-by: Richard Henderson <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

target-sparc: Remove cpu_tmp64 use from softint insns

The use of "tl" functions and a tmp64 is logically incompatible.
Use cpu_tmp0 instead.

Signed-off-by: Richard Henderson <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

target-sparc: Don't use a temporary for gen_dest_fpr_D

In all cases we don't have write-before-read problems.

Signed-off-by: Richard Henderson <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

target-sparc: Remove usage of cpu_tmp64 from most helper functions

Use a locally allocated temporary instead.

Signed-off-by: Richard Henderson <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

target-sparc: Tidy ldfsr, stfsr

Remove the last uses of cpu_tmp32.  Unify the code between sparc64
and sparc32 by using the proper "tl" functions.

Signed-off-by: Richard Henderson <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

target-sparc: Avoid cpu_tmp32 in Write Priv Register

No need to copy to a temporary to store 32 bits.

Signed-off-by: Richard Henderson <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

target-sparc: Avoid cpu_tmp32 in Read Priv Register

We don't need another temporary here.  Load directly into the
register we want to set.

Signed-off-by: Richard Henderson <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

target-sparc: Use get_temp_i32 in gen_dest_fpr_F

Signed-off-by: Richard Henderson <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

target-sparc: Split out get_temp_i32

Signed-off-by: Richard Henderson <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

target-sparc: Make the cpu_addr variable local to load/store handling

Signed-off-by: Richard Henderson <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

target-sparc: Cleanup cpu_src[12] allocation

Now that get_temp_tl is used for get_src[12], we don't need to
pre-allocate these temporaries.

Fallout from this is moving some assignments around cas/casx to
avoid uninitialized variable warnings.

Signed-off-by: Richard Henderson <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

target-sparc: Finish conversion to gen_load_gpr

All users of gen_movl_{reg_TN,TN_reg} are removed.  At the same time,
make cpu_val a local variable for load/store disassembly.

Signed-off-by: Richard Henderson <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

target-sparc: Convert swap to gen_load/store_gpr

Signed-off-by: Richard Henderson <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

target-sparc: Convert asi helpers to gen_*_gpr

Push the DisasContext down so that we can use gen_load/store_gpr
in sode gen_ldda_asi, gen_stda_ast, gen_cas_asi, gen_casx_asi.

Signed-off-by: Richard Henderson <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

target-sparc: Use gen_load_gpr in get_src[12]

This means we can avoid the incoming temporary, though the cleanup
of the existing temporaries is not performed in this patch.

Signed-off-by: Richard Henderson <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

target-sparc: Conversion to gen_*_gpr, part 1

Only handle the easy cases directly within disas_sparc_insn.

Signed-off-by: Richard Henderson <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

target-sparc: Add gen_load/store/dest_gpr

Infrastructure to be used to clean up handling of temporaries.

Signed-off-by: Richard Henderson <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

exec: Make MIN_CODE_GEN_BUFFER_SIZE private to exec.c

It is used nowhere else, and the corresponding MAX_CODE_GEN_BUFFER_SIZE
also lives there.

Signed-off-by: Richard Henderson <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

exec: Allocate code_gen_prologue from code_gen_buffer

We had a hack for arm and sparc, allocating code_gen_prologue to a
special section.  Which, honestly does no good under certain cases.
We've already got limits on code_gen_buffer_size to ensure that all
TBs can use direct branches between themselves; reuse this limit to
ensure the prologue is also reachable.

As a bonus, we get to avoid marking a page of the main executable's
data segment as executable.

Signed-off-by: Richard Henderson <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

exec: Do not use absolute address hints for code_gen_buffer with -fpie

The hard-coded addresses inside alloc_code_gen_buffer only make sense
if we're building an executable that will actually run at the address
we've put into the linker scripts.

When we're building with -fpie, the executable will run at some
random location chosen by the kernel.  We get better placement for
the code_gen_buffer if we allow the kernel to place the memory,
as it will tend to to place it near the executable, based on the
PROT_EXEC bit.

Since code_gen_prologue is always inside the executable, this effect
is easily seen at the end of most TB, with the exit_tb opcode, and
with any calls to helper functions.

Signed-off-by: Richard Henderson <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

exec: Don't make DEFAULT_CODE_GEN_BUFFER_SIZE too large

For ARM we cap the buffer size to 16MB.  Do not allocate 32MB in that case.

Signed-off-by: Richard Henderson <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

exec: Split up and tidy code_gen_buffer

It now consists of:

A macro definition of MAX_CODE_GEN_BUFFER_SIZE with host-specific values,

A function size_code_gen_buffer that applies most of the reasoning for
choosing a buffer size,

Three variations of a function alloc_code_gen_buffer that contain all
of the logic for allocating executable memory via a given allocation
mechanism.

Signed-off-by: Richard Henderson <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

create struct for machine initialization arguments

This should help us to:
- More easily add or remove machine initialization arguments without
  having to change every single machine init function;
- More easily make mechanical changes involving the machine init
  functions in the future;
- Let machine initialization forward the init arguments to other
  functions more easily.

This change was half-mechanical process: first the struct was added with
the local ram_size, boot_device, kernel_*, initrd_*, and cpu_model local
variable initialization to all functions. Then the compiler helped me
locate the local variables that are unused, so they could be removed.

Signed-off-by: Blue Swirl <[email protected]>

vga: remove CONFIG_BOCHS_VBE

Signed-off-by: Gerd Hoffmann <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

vga: add specs for standard vga

Signed-off-by: Gerd Hoffmann <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

vga: add mmio bar to standard vga

This patch adds a mmio bar to the qemu standard vga which allows to
access the standard vga registers and bochs dispi interface registers
via mmio.

Cc: Benjamin Herrenschmidt <[email protected]>
Signed-off-by: Gerd Hoffmann <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

vga: fix indention

Signed-off-by: Gerd Hoffmann <[email protected]>
Signed-off-by: Blue Swirl <[email protected]>

tests/tcg: fix build

This broke when the tests were moved from tests/ to tests/tcg/.

On x86_64 host/i386-linux-user non-kvm guest, test-i386 and test-mmap are broken, but at least they build.

To build/run the tests:
$ cd $BUILD_PATH/tests/tcg
$ SRC_PATH=path/to/qemu make <target>

Signed-off-by: Catalin Patulea <[email protected]>
Reviewed-by: Peter Maydell <[email protected]>
Signed-off-by: Aurelien Jarno <[email protected]>

allow make {dist, }clean work w/out configure

There's no reason to require configure to run before running a clean
target, so check MAKECMDGOALS before.

Signed-off-by: Mike Frysinger <[email protected]>
Signed-off-by: Aurelien Jarno <[email protected]>

tci: fix build breakage for target-sparc

commit c28ae41 introduced GETPC() usage for sparc, which is currently
not defined when building with --enable-tcg-interpreter. Add sparc to
the list of targets we selectively define GETPC() for.

Signed-off-by: Michael Roth <[email protected]>
Reviewed-by: Stefan Weil <[email protected]>
Signed-off-by: Aurelien Jarno <[email protected]>

MAINTAINERS: Update email address for Stefan Hajnoczi

Switch to my new work email address from which I am contributing.

Signed-off-by: Stefan Hajnoczi <[email protected]>
Signed-off-by: Aurelien Jarno <[email protected]>

configure: Fix CONFIG_QEMU_HELPERDIR generation

We need to evaluate $libexecdir in configure, otherwise we literally end
up with "${prefix}/libexec" instead of the absolute path as
CONFIG_QEMU_HELPERDIR.

Signed-off-by: Jan Kiszka <[email protected]>
Signed-off-by: Aurelien Jarno <[email protected]>

qemu-options.hx: Change from recommending '?' to 'help'

Update the -help output and documentation so that it recommends
'help' rather than '?' for the various "list valid values for this
option" cases. '?' is deprecated (as it can fail confusingly if
not quoted), so it's better to steer users towards 'help'. ('?'
still works, for backwards compatibility.)

This is the -help option part of the change otherwise done in
commit c8057f9, since we are now past release 1.2 and free to
change our help text without worrying about breaking libvirt.

Signed-off-by: Peter Maydell <[email protected]>
Reviewed-by: Eric Blake <[email protected]>
Signed-off-by: Aurelien Jarno <[email protected]>

Merge branch 'linux-user-for-upstream' of git://git.linaro.org/people/rikuvoipio/qemu

* 'linux-user-for-upstream' of git://git.linaro.org/people/rikuvoipio/qemu:
  linux-user: register align p{read, write}64
  linux-user: ppc: mark as long long aligned
  tcg: Remove TCG_TARGET_HAS_GUEST_BASE define
  configure: Remove unnecessary host_guest_base code
  linux-user: If loading fails, print error as string, not number
  linux-user: Fix siginfo handling
  alpha-linux-user: Fix sigaltstack structure definition
  linux-user: Implement gethostname
  linux-user: Perform more checks on iovec lists
  linux-user: fix multi-threaded /proc/self/maps
  linux-user: fix statfs

Merge branch 'trivial-patches' of git://github.com/stefanha/qemu

* 'trivial-patches' of git://github.com/stefanha/qemu:
  ui/vnc-jobs.c: Fix minor typos in comments
  net/tap-win32: Fix compiler warning caused by missing include statement
  configure: Remove unused parameters from main function
  target-arm/neon_helper: Remove obsolete FIXME comment
  targphys.h: Don't define target_phys_addr_t for user-mode emulators
  ui/vnc: Only report/use TIGHT_PNG encoding if enabled.

microblaze: Update PC before simulating syscall

Fixes a clone() emulation bug were the new thread starts
at the point of the syscall and thus clones in a loop.

Signed-off-by: Edgar E. Iglesias <[email protected]>

microblaze: Support setting of TLS ptr

Signed-off-by: Edgar E. Iglesias <[email protected]>

ui/vnc-jobs.c: Fix minor typos in comments

Fix some minor typos/grammar errors in comments.

Signed-off-by: Peter Maydell <[email protected]>
Reviewed-by: Stefan Weil <[email protected]>
Signed-off-by: Stefan Hajnoczi <[email protected]>

net/tap-win32: Fix compiler warning caused by missing include statement

The include file for net_init_tap was missing:

net/tap-win32.c:703:
 warning: no previous prototype for ‘net_init_tap’

Signed-off-by: Stefan Weil <[email protected]>
Signed-off-by: Stefan Hajnoczi <[email protected]>

configure: Remove unused parameters from main function

This modification is required if compiler option -Wunused-parameter is activated.

Signed-off-by: Stefan Weil <[email protected]>
Signed-off-by: Stefan Hajnoczi <[email protected]>

target-arm/neon_helper: Remove obsolete FIXME comment

Commit 33ebc29 fixed the bugs in the implementation of VQRSHL,
but forgot to remove the FIXME comment...

Signed-off-by: Peter Maydell <[email protected]>
Signed-off-by: Stefan Hajnoczi <[email protected]>

targphys.h: Don't define target_phys_addr_t for user-mode emulators

Commit 4be403c accidentally defined the target_phys_addr_t type when
building user-mode emulators. Since the type doesn't really make
any sense except for system emulators, avoid defining it when building
in user mode.

Signed-off-by: Peter Maydell <[email protected]>
Acked-by: Andreas Färber <[email protected]>
Signed-off-by: Stefan Hajnoczi <[email protected]>

ui/vnc: Only report/use TIGHT_PNG encoding if enabled.

If TIGHT_PNG is not enabled by the --enable-vnc-png configure flag
then do not report to the client that it is supported.

Also, since TIGHT_PNG is the same as the TIGHT encoding but with the
filter/copy replaced with PNG data, adding it to the supported
encodings list when it is disabled will cause the TIGHT encoding to be
used even though the client requested TIGHT_PNG.

Signed-off-by: Joel Martin <[email protected]>
Signed-off-by: Stefan Hajnoczi <[email protected]>

target-arm/translate: Fix RRX operands

Instructions that both use the RRX second operand and update CS were
incorrect, as the Carry flag was updated too early. An example of such an
instruction would be:

ands r12,r13,RRX

Ands, because of the "s" flag will update the carry flag. But the RRX second
operand rotates through the C flag which should happen before the update.
Fixed the ordering of the two, the old carry is read by "r13,RRX" before being
updated.

Signed-off-by: Peter Crosthwaite <[email protected]>
Reported-by: Vinesh Peringat <[email protected]>
Reviewed-by: Peter Maydell <[email protected]>
Signed-off-by: Aurelien Jarno <[email protected]>

vfio-pci: Mark non-migratable

We haven't magically fixed this yet.  Toss in a description too.

Signed-off-by: Alex Williamson <[email protected]>

vfio-pci: Fix debug build

Stray variable from before MSI-X rework

Signed-off-by: Alex Williamson <[email protected]>

cpus: create qemu_in_vcpu_thread()

Old code used !io_thread to know if a thread was an vcpu or not.  That
fails when we introduce the iothread.

Signed-off-by: Juan Quintela <[email protected]>

savevm: make qemu_file_put_notify() return errors

Signed-off-by: Juan Quintela <[email protected]>
Reviewed-by: Paolo Bonzini <[email protected]>

savevm: un-export qemu_file_set_error()

Signed-off-by: Juan Quintela <[email protected]>
Reviewed-by: Paolo Bonzini <[email protected]>

block-migration: handle errors with the return codes correctly

Signed-off-by: Juan Quintela <[email protected]>
Reviewed-by: Paolo Bonzini <[email protected]>

block-migration:  Switch meaning of return value

Make consistent the result of blk_mig_save_dirty_block() and
mig_save_device_dirty()

Signed-off-by: Juan Quintela <[email protected]>
Reviewed-by: Paolo Bonzini <[email protected]>

block-migration: make flush_blks() return errors

This means we don't need to pass through qemu_file to get the errors.
Adjust all callers.

Signed-off-by: Juan Quintela <[email protected]>
Reviewed-by: Paolo Bonzini <[email protected]>

buffered_file: buffered_put_buffer() don't need to set last_error

Callers on savevm.c:qemu_fflush() will set it.

Signed-off-by: Juan Quintela <[email protected]>
Reviewed-by: Paolo Bonzini <[email protected]>