qemu-nbd: drop loop which can never loop

For some reason nbd_client_thread() has a do..while loop which can never
loop, the condition is bogus because we would take a goto instead.  Drop
the loop.

Reported-by: Dr David Alan Gilbert <[email protected]>
Reviewed-by: Paolo Bonzini <[email protected]>
Signed-off-by: Stefan Hajnoczi <[email protected]>

Make python mandatory

The QEMU build depends on Python so make it an explicit requirement.

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

net/socket.c: Fix fd leak in net_socket_listen_init() error paths

Fix a leak of a file descriptor due to missing closesocket() calls
in error paths in net_socket_listen_init().

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

gdbstub: Fix fd leak in gdbserver_open() error path

Fix a leak of a file descriptor in error exit paths in
gdbserver_open().

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

configure: Fix test for supported host CPU type

The test for whether the host CPU is supported had several problems:
 * the attempt to fall back to TCI was done as a duplicate
   test, very late (so "--cpu foo" would fail early but "--cpu unicore32"
   would fail late, differently, and after configure had already
   printed a lot of output)
 * a number of CPUs only supported as guests were included in the
   list of CPUs we would accept as valid hosts, which would result
   in a late compile failure on those systems rather than a
   configure failure or fallback to TCI
 * bailing out for an unsupported CPU happened before the main
   option parsing, so "configure --help" wouldn't work

Fix these by folding the setting of ARCH into the first test for
supported host CPU, removing spurious guest-only CPU names from it,
and moving the "fall back to TCI" code earlier.

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

configure: CONFIG_QEMU_INTERP_PREFIX only for user mode

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

scsi virtio-blk usb-msd: Clean up device init error messages

Replace

    error_report("DEVICE-NAME: MESSAGE");

by just

    error_report("MESSAGE");

in block device init functions.

DEVICE-NAME is bogus in some cases: it's "scsi-disk" for device
scsi-hd and scsi-cd, "virtio-blk-pci" for virtio-blk-s390, and
"usb-msd" for usb-storage.

There is no real need to put a device name in the message, because
error_report() points to the offending command line option already:

$ qemu-system-x86_64 --nodefaults --enable-kvm -vnc :0 -S -monitor stdio -usb -device virtio-blk-pci
upstream-qemu: -device virtio-blk-pci: virtio-blk-pci: drive property not set
upstream-qemu: -device virtio-blk-pci: Device 'virtio-blk-pci' could not be initialized

And for a monitor command, it's obvious anyway:

$ qemu-system-x86_64 --nodefaults --enable-kvm -vnc :0 -S -monitor stdio -usb
(qemu) device_add virtio-blk-pci
virtio-blk-pci: drive property not set
Device 'virtio-blk-pci' could not be initialized

Reported-by: Amit Shah <[email protected]>
Signed-off-by: Markus Armbruster <[email protected]>
Signed-off-by: Stefan Hajnoczi <[email protected]>

Strip trailing '\n' from error_report()'s first argument (again)

Commit 6daf194d got rid of them, but Hans and Gerd added some more
lately.  Tracked down with this Coccinelle semantic patch:

@r@
    expression fmt;
    position p;
@@
    error_report(fmt, ...)@p
@script:python@
    fmt << r.fmt;
    p << r.p;
@@
if "\\n" in str(fmt):
    print "%s:%s:%s:%s" % (p[0].file, p[0].line, p[0].column, fmt)

Signed-off-by: Markus Armbruster <[email protected]>

qemu-options.hx: fix tls-channel help text

Remove the default compiled out tunnel channel, add the always available
cursor channel. Optimally the man page would depend on compiled in
options, but that's harder to do.

RHBZ: 688586

Acked-by: Gerd Hoffmann <[email protected]>
Signed-off-by: Alon Levy <[email protected]>
Signed-off-by: Stefan Hajnoczi <[email protected]>

Merge remote-tracking branch 'pmaydell/arm-devs.for-upstream' into staging

* pmaydell/arm-devs.for-upstream:
  add L2x0/PL310 cache controller device
  arm: add dummy gic security registers
  arm: Set frequencies for arm_timer
  arm: add missing scu registers
  hw/omap_gpmc: Fix region map/unmap when configuring prefetch engine
  hw/omap1.c: Drop unused includes
  hw/omap1.c: Separate dpll_ctl from omap_mpu_state
  hw/omap1.c: Separate PWT from omap_mpu_state
  hw/omap1.c: Separate PWL from omap_mpu_state
  hw/omap1.c: omap_mpuio_init() need not be public
  hw/pl110.c: Add post-load hook to invalidate display
  hw/pl181.c: Add save/load support

add L2x0/PL310 cache controller device

This is just a dummy device for ARM L2 cache controllers, based on the
pl310. The cache type parameter can be defined by a property value
and has a meaningful default.

Signed-off-by: Rob Herring <[email protected]>
Signed-off-by: Mark Langsdorf <[email protected]>
[Peter Maydell: removed stray blank line at end]
Signed-off-by: Peter Maydell <[email protected]>

arm: add dummy gic security registers

Implement handling for the RAZ/WI gic security registers.

Signed-off-by: Rob Herring <[email protected]>
Signed-off-by: Mark Langsdorf <[email protected]>
Signed-off-by: Peter Maydell <[email protected]>

arm: Set frequencies for arm_timer

Use qdev properties to allow board modelers to set the frequencies
for the sp804 timer. Each of the sp804's timers can have an
individual frequency. The timers default to 1MHz.

Signed-off-by: Mark Langsdorf <[email protected]>
Reviewed-by: Andreas Färber <[email protected]>
Signed-off-by: Peter Maydell <[email protected]>

arm: add missing scu registers

Add power control register to a9mpcore

Signed-off-by: Rob Herring <[email protected]>
Signed-off-by: Mark Langsdorf <[email protected]>
Signed-off-by: Peter Maydell <[email protected]>

hw/omap_gpmc: Fix region map/unmap when configuring prefetch engine

When configuring the prefetch engine (and also when resetting from
a state where the prefetch engine was enabled) be careful to adhere
to the "unmap/change config fields/map" ordering, to avoid trying
to delete the wrong MemoryRegions. This fixes an assertion failure
in some cases.

Signed-off-by: Peter Maydell <[email protected]>
Reported-by: Alexander Graf <[email protected]>
Tested-by: Alexander Graf <[email protected]>

hw/omap1.c: Drop unused includes

Drop includes of qemu-timer.h, qemu-char.h and pc.h as they are no
longer needed.

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

hw/omap1.c: Separate dpll_ctl from omap_mpu_state

Signed-off-by: Juha Riihimäki <[email protected]>
[Riku Voipio: Fixes and restructuring patchset]
Signed-off-by: Riku Voipio <[email protected]>
[Peter Maydell: More fixes and cleanups for upstream submission]
Signed-off-by: Peter Maydell <[email protected]>

hw/omap1.c: Separate PWT from omap_mpu_state

Signed-off-by: Juha Riihimäki <[email protected]>
[Riku Voipio: Fixes and restructuring patchset]
Signed-off-by: Riku Voipio <[email protected]>
[Peter Maydell: More fixes and cleanups for upstream submission]
Signed-off-by: Peter Maydell <[email protected]>

Remove IO_MEM_SHIFT

We no longer use any of the lower bits of a ram_addr, so we might as well
use them for the io table index.  This increases the number of potential
I/O handlers by a factor of 8.

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

Drop IO_MEM_ROMD

Unlike ->readonly, ->readable is not inherited from aliase, so we can simply
query the memory region.

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

Remove IO_MEM_SUBPAGE

Replace with a MemoryRegion flag.

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

Direct dispatch through MemoryRegion

Now that all mmio goes through MemoryRegions, we can convert
io_mem_opaque to be a MemoryRegion pointer, and remove the thunks
that convert from old-style CPU{Read,Write}MemoryFunc to MemoryRegionOps.

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

Convert io_mem_watch to be a MemoryRegion

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

Convert IO_MEM_SUBPAGE_RAM to be a MemoryRegion

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

Convert the subpage wrapper to be a MemoryRegion

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

Switch cpu_register_physical_memory_log() to use MemoryRegions

Still internally using ram_addr.

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

Convert IO_MEM_{RAM,ROM,UNASSIGNED,NOTDIRTY} to MemoryRegions

Convert the fixed-address IO_MEM_RAM, IO_MEM_ROM, IO_MEM_UNASSIGNED,
and IO_MEM_NOTDIRTY io handlers to MemoryRegions.  These aren't real
regions, since they are never added to the memory hierarchy, but they
allow reuse of the dispatch functionality.

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

Uninline get_page_addr_code()

Its use of IO_MEM_ROM and friends will later cause #include loops; and it
is too large to merit inlining.

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

Avoid range comparisons on io index types

The code sometimes uses range comparisons on io indexes (e.g.
index =< IO_MEM_ROM).  Avoid these as they make moving to objects harder.

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

Fix wrong region_offset when overlaying a page with another

cpu_register_physical_memory_log() does not update region_offset
if a page was previously registered for the same address.  This
could cause mmio accesses going to the wrong place, by using the
old region_offset.

Signed-off-by: Avi Kivity <[email protected]>
Acked-by: Andreas Färber <[email protected]>
Reviewed-by: Richard Henderson <[email protected]>

memory: remove MemoryRegion::backend_registered

backend_registered was used to lazify the process of registering an
mmio region, since the it is different for the I/O address space and
the memory address space.  However, it also makes registration dependent
on the region being visible in the address space.  This is not the case
for "fake" regions, like watchpoints or IO_MEM_UNASSIGNED.

Remove backend_registered and always initialize the region.  If it turns
out to be part of the I/O address space, we've wasted an I/O slot, but
that's not too bad.  In any case this will be optimized later on.

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

memory: move mmio access to functions

Currently mmio access goes directly to the io_mem_{read,write} arrays.
In preparation for eliminating them, add indirection via a function.

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

exec: make phys_page_find() return a temporary

Instead of returning a PhysPageDesc pointer, return a temporary.
This lets us move away from actually storing PhysPageDesc's, and
instead sythesising them when needed.

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

memory: move endianness compensation to memory core

Instead of doing device endianness compensation in cpu_register_io_memory(),
do it in the memory core.

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

memory: obsolete more dirty memory related functions

No longer used outside memory.c and exec.c.

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

xen: convert framebuffer dirty tracking to memory API

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

memory: obsolete cpu_physical_memory_[gs]et_dirty_tracking()

The getter is no longer used, so it is completely removed.

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

Convert ram_load() to the memory API

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

Remove support for version 3 ram_load

Version 3 ram_load depends on ram_addrs, which are not stable.  Version 4
was introduced in 0.13 (and RHEL 6), so this means live migration from 0.12
and earlier to 1.1 or later will not work.

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

Sort RAMBlocks by ID for migration, not by ram_addr

ram_addr is (a) unstable (b) going away.  Sort by idstr instead.

Commit b2e0a138e initially introduced the sorting for the purpose
of improving debuggability.  After this patch, the order is still
stable, but perhaps less usable by a human.

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

Switch ram_save to the memory API

Avoid using ram_addr_t, instead use (MemoryRegion *, offset) pairs.

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

Store MemoryRegion in RAMBlock

As a step in moving live migration from RAMBlocks to MemoryRegions,
store the MemoryRegion in a RAMBlock.

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

vmstate, memory: decouple vmstate from memory API

Currently creating a memory region automatically registers it for
live migration.  This differs from other state (which is enumerated
in a VMStateDescription structure) and ties the live migration code
into the memory core.

Decouple the two by introducing a separate API, vmstate_register_ram(),
for registering a RAM block for migration.  Currently the same
implementation is reused, but later it can be moved into a separate list,
and registrations can be moved to VMStateDescription blocks.

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

memory: introduce memory_region_name()

Trivial accessor for the name attribute.

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

hw/omap1.c: Separate PWL from omap_mpu_state

Signed-off-by: Juha Riihimäki <[email protected]>
[Riku Voipio: Fixes and restructuring patchset]
Signed-off-by: Riku Voipio <[email protected]>
[Peter Maydell: More fixes and cleanups for upstream submission]
Signed-off-by: Peter Maydell <[email protected]>

hw/omap1.c: omap_mpuio_init() need not be public

omap_mpuio_init() is only used and defined in omap1.c, so make it static.

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

hw/pl110.c: Add post-load hook to invalidate display

Add a post-load hook which invalidates the display. In particular, if we
don't do this and the display size we've just reloaded is larger than
the default then we will segfault trying to read off the end of the buffer.

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

hw/pl181.c: Add save/load support

Add save/load support to the PL181.

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

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

* qemu-kvm/memory/page_desc: (22 commits)
  Remove cpu_get_physical_page_desc()
  sparc: avoid cpu_get_physical_page_desc()
  virtio-balloon: avoid cpu_get_physical_page_desc()
  vhost: avoid cpu_get_physical_page_desc()
  kvm: avoid cpu_get_physical_page_desc()
  memory: remove CPUPhysMemoryClient
  xen: convert to MemoryListener API
  memory: temporarily add memory_region_get_ram_addr()
  xen, vga: add API for registering the framebuffer
  vhost: convert to MemoryListener API
  kvm: convert to MemoryListener API
  kvm: switch kvm slots to use host virtual address instead of ram_addr_t
  memory: add API for observing updates to the physical memory map
  memory: replace cpu_physical_sync_dirty_bitmap() with a memory API
  framebuffer: drop use of cpu_physical_sync_dirty_bitmap()
  loader: remove calls to cpu_get_physical_page_desc()
  framebuffer: drop use of cpu_get_physical_page_desc()
  memory: introduce memory_region_find()
  memory: add memory_region_is_logging()
  memory: add memory_region_is_rom()
  ...

Remove cpu_get_physical_page_desc()

No longer used.

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

sparc: avoid cpu_get_physical_page_desc()

This reaches into the innards of the memory core, which are being
changed.  Switch to a memory API version.

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

virtio-balloon: avoid cpu_get_physical_page_desc()

This reaches into the innards of the memory core, which are being
changed.  Switch to a memory API version.

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

vhost: avoid cpu_get_physical_page_desc()

This reaches into the innards of the memory core, which are being
changed.  Switch to a memory API version.

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

kvm: avoid cpu_get_physical_page_desc()

This reaches into the innards of the memory core, which are being
changed.  Switch to a memory API version.

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

memory: remove CPUPhysMemoryClient

No longer used.

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

xen: convert to MemoryListener API

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

Fix qapi code generation wrt parallel build

Make's multiple output syntax

  x.c x.h: x.template
       gen < x.template

actually invokes the command once for x.c and once for x.h (with differing $@
in each invocation).  During a parallel build, the two commands may be invoked
in parallel; this opens up a race, where the second invocation trashes a file
supposedly produced during the first, and now in use by a dependent command.

The various qapi code generators are susceptible to this; fix by making them
generate just one file per invocation.

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

Merge remote-tracking branch 'aneesh/for-upstream' into staging

* aneesh/for-upstream:
  scripts/analyse-9p-simpletrace.py: Add symbolic names for 9p operations.
  hw/9pfs: iattr_valid flags are kernel internal flags map them to 9p values.
  hw/9pfs: Use the correct signed type for different variables
  hw/9pfs: replace iovec manipulation with QEMUIOVector

Merge remote-tracking branch 'bonzini/nbd-for-anthony' into staging

* bonzini/nbd-for-anthony: (26 commits)
  nbd: add myself as maintainer
  qemu-nbd: throttle requests
  qemu-nbd: asynchronous operation
  qemu-nbd: add client pointer to NBDRequest
  qemu-nbd: move client handling to nbd.c
  qemu-nbd: use common main loop
  link the main loop and its dependencies into the tools
  qemu-nbd: introduce NBDRequest
  qemu-nbd: introduce NBDExport
  qemu-nbd: introduce nbd_do_receive_request
  qemu-nbd: more robust handling of invalid requests
  qemu-nbd: introduce nbd_do_send_reply
  qemu-nbd: simplify nbd_trip
  move corking functions to osdep.c
  qemu-nbd: remove data_size argument to nbd_trip
  qemu-nbd: remove offset argument to nbd_trip
  Update ioctl order in nbd_init() to detect EBUSY
  nbd: add support for NBD_CMD_TRIM
  nbd: add support for NBD_CMD_FLUSH
  nbd: add support for NBD_CMD_FLAG_FUA
  ...

enable architectural PMU cpuid leaf for kvm

Signed-off-by: Gleb Natapov <[email protected]>
Signed-off-by: Marcelo Tosatti <[email protected]>

Set numa topology for max_cpus

qemu-kvm passes numa/SRAT topology information for smp_cpus to SeaBIOS. However
SeaBIOS always expects to setup max_cpus number of SRAT cpu entries
(MaxCountCPUs variable in build_srat function of Seabios). When qemu-kvm runs
with smp_cpus != max_cpus (e.g. -smp 2,maxcpus=4), Seabios will mistakenly use
memory SRAT info for setting up CPU SRAT entries for the offline CPUs. Wrong
SRAT memory entries are also created. This breaks NUMA in a guest.
Fix by setting up SRAT info for max_cpus in qemu-kvm.

Signed-off-by: Vasilis Liaskovitis <[email protected]>
Signed-off-by: Marcelo Tosatti <[email protected]>

kvm: x86: Drop redundant apic base and tpr update from kvm_get_sregs

The latter was already commented out, the former is redundant as well.
We always get the latest changes after return from the guest via
kvm_arch_post_run.

Signed-off-by: Jan Kiszka <[email protected]>
Signed-off-by: Marcelo Tosatti <[email protected]>

kvm: x86: Avoid runtime allocation of xsave buffer

Keep a per-VCPU xsave buffer for kvm_put/get_xsave instead of
continuously allocating and freeing it on state sync.

Signed-off-by: Jan Kiszka <[email protected]>
Signed-off-by: Marcelo Tosatti <[email protected]>

kvm: x86: Use symbols for all xsave field

Field 0 (FCW+FSW) and 1 (FTW+FOP) were hard-coded so far.

Signed-off-by: Jan Kiszka <[email protected]>
Signed-off-by: Marcelo Tosatti <[email protected]>

nbd: add myself as maintainer

Not planning to do much else, hence listing it as "Odd Fixes".

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

qemu-nbd: throttle requests

Limiting the number of in-flight requests is implemented very simply
with a can_read callback.  It does not require a semaphore, unlike the
client side in block/nbd.c, because we can throttle directly the creation
of coroutines.  The client side can have a coroutine created at any time
when an I/O request is made.

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

qemu-nbd: asynchronous operation

Using coroutines enable asynchronous operation on both the network and
the block side.  Network can be owned by two coroutines at the same time,
one writing and one reading.  On the send side, mutual exclusion is
guaranteed by a CoMutex.  On the receive side, mutual exclusion is
guaranteed because new coroutines immediately start receiving data,
and no new coroutines are created as long as the previous one is receiving.

Between receive and send, qemu-nbd can have an arbitrary number of
in-flight block transfers.  Throttling is implemented by the next
patch.

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

qemu-nbd: add client pointer to NBDRequest

By attaching a client to an NBDRequest, we can avoid passing around the
socket descriptor and data buffer.

Also, we can now manage the reference count for the client in
nbd_request_get/put request instead of having to do it ourselved in
nbd_read.  This simplifies things when coroutines are used.

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

qemu-nbd: move client handling to nbd.c

This patch sets up the fd handler in nbd.c instead of qemu-nbd.c.  It
introduces NBDClient, which wraps the arguments to nbd_trip in a single
structure, so that we can add a notifier to it.  This way, qemu-nbd can
know about disconnections.

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

qemu-nbd: use common main loop

Using a single main loop for sockets will help yielding from the socket
coroutine back to the main loop, and later reentering it.

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

link the main loop and its dependencies into the tools

Using the main loop code from QEMU enables tools to operate fully
asynchronously.  Advantages include better Windows portability (for some
definition of portability) over glib's.

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

qemu-nbd: introduce NBDRequest

Move the buffer from NBDExport to a new structure, so that it will be
possible to have multiple in-flight requests for the same export
(and for the same client too---we get that for free).

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

qemu-nbd: introduce NBDExport

Wrap the common parameters of nbd_trip and nbd_negotiate in a
single opaque struct.

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

qemu-nbd: introduce nbd_do_receive_request

Group the receiving of a response and the associated data into a new function.

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

qemu-nbd: more robust handling of invalid requests

Fail invalid requests with EINVAL instead of dropping them into
the void.

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

qemu-nbd: introduce nbd_do_send_reply

Group the sending of a reply and the associated data into a new function.
Without corking, the caller would be forced to leave 12 free bytes at the
beginning of the data pointer.  Not too ugly, but still ugly. :)

Using nbd_do_send_reply everywhere will help when the routine will set up
the write handler that re-enters the send coroutine.

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

qemu-nbd: simplify nbd_trip

Use TCP_CORK to remove a violation of encapsulation, that would later
require nbd_trip to know too much about an NBD reply.

We could also switch to sendmsg (qemu_co_sendv) later, it is even
easier once coroutines are in.

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

move corking functions to osdep.c

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

qemu-nbd: remove data_size argument to nbd_trip

The size of the buffer is in practice part of the protocol.

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

qemu-nbd: remove offset argument to nbd_trip

The argument is write-only.

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

Update ioctl order in nbd_init() to detect EBUSY

Update ioctl(s) in nbd_init() to detect device busy early.

Current nbd_init() issues NBD_CLEAR_SOCKET before NBD_SET_SOCKET, if issuing
"qemu-nbd -c /dev/nbd0 disk.img" twice, the second time won't detect EBUSY in
nbd_init(), but in nbd_client will report EBUSY and do clear socket (the 1st
time command will be affacted too because of no socket any more.)

No change to previous version.

Signed-off-by: Chunyan Liu <[email protected]>
Signed-off-by: Paolo Bonzini <[email protected]>

nbd: add support for NBD_CMD_TRIM

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

nbd: add support for NBD_CMD_FLUSH

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

nbd: add support for NBD_CMD_FLAG_FUA

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

nbd: fix error handling in the server

bdrv_read and bdrv_write return negative errno values, not -1.

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

nbd: allow multiple in-flight requests

Allow sending up to 16 requests, and drive the replies to the coroutine
that did the request.  The code is written to be exactly the same as
before this patch when MAX_NBD_REQUESTS == 1 (modulo the extra mutex
and state).

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

nbd: split requests

qemu-nbd has a limit of slightly less than 1M per request.  Work
around this in the nbd block driver.

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

nbd: switch to asynchronous operation

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

sheepdog: move coroutine send/recv function to generic code

Outside coroutines, avoid busy waiting on EAGAIN by temporarily
making the socket blocking.

The API of qemu_recvv/qemu_sendv is slightly different from
do_readv/do_writev because they do not handle coroutines.  It
returns the number of bytes written before encountering an
EAGAIN.  The specificity of yielding on EAGAIN is entirely in
qemu-coroutine.c.

Reviewed-by: MORITA Kazutaka <[email protected]>
Signed-off-by: Paolo Bonzini <[email protected]>

virtio-serial-bus: Ports are expected to implement 'have_data' callback

There's no need to check if ports can accept any incoming data from the
guest each time the guest sends data.  Check if the port implements such
functionality during port initialisation.

Signed-off-by: Amit Shah <[email protected]>
Signed-off-by: Anthony Liguori <[email protected]>

virtio-console: Properly initialise class methods

The earlier code really was a hack: initialising class methods in an
object init function as noted by Anthony.

The motivation for that was to not have the virtio-serial-bus call into
the callback functions if there was no chardev backend registered.
However, that really wasn't a worthwhile optimisation, and definitely
not one that was well-implemented.  Get rid of it.

Reported-by: Anthony Liguori <[email protected]>
Signed-off-by: Amit Shah <[email protected]>
Signed-off-by: Anthony Liguori <[email protected]>

virtio-console: Check if chardev backends available before calling into them

For the callback functions invoked by the virtio-serial-bus code, check
if we have chardev backends registered before we call into the chardev
functions.

Signed-off-by: Amit Shah <[email protected]>
Reported-by: Anthony Liguori <[email protected]>
Signed-off-by: Amit Shah <[email protected]>
Signed-off-by: Anthony Liguori <[email protected]>

add qemu_send_full and qemu_recv_full

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

scripts/analyse-9p-simpletrace.py: Add symbolic names for 9p operations.

Currently, we just print the numerical value of 9p operation identifier in
case of RERROR which is less meaningful for readability. Mapping 9p
operation ids to symbolic names provides a better tracelog:

RERROR (tag = 1 , id = TWALK , err = " No such file or directory ")
RERROR (tag = 1 , id = TUNLINKAT , err = " Directory not empty ")

This patch provides a dictionary of all possible 9p operation symbols mapped
to their numerical identifiers which are likely to be used in future at
various places in this script.

Signed-off-by: Harsh Prateek Bora <[email protected]>
Signed-off-by: Aneesh Kumar K.V <[email protected]>

hw/9pfs: iattr_valid flags are kernel internal flags map them to 9p values.

Kernel internal values can change, add protocol values for these constant and
use them.

Signed-off-by: Aneesh Kumar K.V <[email protected]>

hw/9pfs: Use the correct signed type for different variables

Signed-off-by: Aneesh Kumar K.V <[email protected]>

hw/9pfs: replace iovec manipulation with QEMUIOVector

The v9fs_read() and v9fs_write() functions rely on iovec[] manipulation
code should be replaced with QEMUIOVector to avoid duplicating code.
In the future it may be possible to make the code even more concise by
using QEMUIOVector consistently across virtio and 9pfs.

The "v" format specifier for pdu_marshal() and pdu_unmarshal() is
dropped since it does not actually pack/unpack anything.  The specifier
was also not implemented to update the offset variable and could only be
used at the end of a format string, another sign that this shouldn't
really be a format specifier.  Instead, see the new
v9fs_init_qiov_from_pdu() function.

This change avoids a possible iovec[] buffer overflow when indirect
vrings are used since the number of vectors is now limited by the
underlying VirtQueueElement and cannot be out-of-bounds.

Signed-off-by: Stefan Hajnoczi <[email protected]>
Signed-off-by: Aneesh Kumar K.V <[email protected]>

sd: Remember to reset .expecting_acmd on reset.

Signed-off-by: Andrzej Zaborowski <[email protected]>

hw/sd.c: Clear status bits when read via response r6

Response format r6 includes a subset of the status bits;
clear the clear-on-read bits which are read by an r6 response.

Signed-off-by: Peter Maydell <[email protected]>
Signed-off-by: Andrzej Zaborowski <[email protected]>

hw/sd.c: Correct handling of APP_CMD status bit

Fix some bugs in our implementation of the APP_CMD status bit:
 * the response to an ACMD should have APP_CMD set, not cleared
 * if an illegal ACMD is sent then the next command should be
   handled as a normal command

This requires that we split "card is expecting an ACMD" from
the state of the APP_CMD status bit (the latter indicates
both "expecting ACMD" and "that was an ACMD").

Signed-off-by: Peter Maydell <[email protected]>
Signed-off-by: Andrzej Zaborowski <[email protected]>