The pseries machines introduced the concept of 'unplug timeout' for CPU
hotunplugs. The idea was to circunvent a deficiency in the pSeries
specification (PAPR), that currently does not define a proper way for
the hotunplug to fail. If the guest refuses to release the CPU (see [1]
for an example) there is no way for QEMU to detect the failure.
Further discussions about how to send a QAPI event to inform about the
hotunplug timeout [2] exposed problems that weren't predicted back when
the idea was developed. Other QEMU machines don't have any type of
hotunplug timeout mechanism for any device, e.g. ACPI based machines
have a way to make hotunplug errors visible to the hypervisor. This
would make this timeout mechanism exclusive to pSeries, which is not
ideal.
The real problem is that a QAPI event that reports hotunplug timeouts
puts the management layer (namely Libvirt) in a weird spot. We're not
telling that the hotunplug failed, because we can't be 100% sure of
that, and yet we're resetting the unplug state back, preventing any
DEVICE_DEL events to reach out in case the guest decides to release the
device. Libvirt would need to inspect the guest itself to see if the
device was released or not, otherwise the internal domain states will be
inconsistent. Moreover, Libvirt already has an 'unplug timeout'
concept, and a QEMU side timeout would need to be juggled together with
the existing Libvirt timeout.
All this considered, this solution ended up creating more trouble than
it solved. This patch reverts the 3 commits that introduced the timeout
mechanism for CPU hotplugs in pSeries machines.
This reverts commit 4515a5f786
"qemu_timer.c: add timer_deadline_ms() helper"
This reverts commit d1c2e3ce3d
"spapr_drc.c: add hotunplug timeout for CPUs"
This reverts commit 51254ffb32
"spapr_drc.c: introduce unplug_timeout_timer"
[1] https://bugzilla.redhat.com/show_bug.cgi?id=1911414
[2] https://lists.gnu.org/archive/html/qemu-devel/2021-03/msg04682.html
CC: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Daniel Henrique Barboza <danielhb413@gmail.com>
Message-Id: <20210401000437.131140-2-danielhb413@gmail.com>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
There is a reliable way to make a CPU hotunplug fail in the pseries
machine. Hotplug a CPU A, then offline all other CPUs inside the guest
but A. When trying to hotunplug A the guest kernel will refuse to do it,
because A is now the last online CPU of the guest. PAPR has no 'error
callback' in this situation to report back to the platform, so the guest
kernel will deny the unplug in silent and QEMU will never know what
happened. The unplug pending state of A will remain until the guest is
shutdown or rebooted.
Previous attempts of fixing it (see [1] and [2]) were aimed at trying to
mitigate the effects of the problem. In [1] we were trying to guess
which guest CPUs were online to forbid hotunplug of the last online CPU
in the QEMU layer, avoiding the scenario described above because QEMU is
now failing in behalf of the guest. This is not robust because the last
online CPU of the guest can change while we're in the middle of the
unplug process, and our initial assumptions are now invalid. In [2] we
were accepting that our unplug process is uncertain and the user should
be allowed to spam the IRQ hotunplug queue of the guest in case the CPU
hotunplug fails.
This patch presents another alternative, using the timeout
infrastructure introduced in the previous patch. CPU hotunplugs in the
pSeries machine will now timeout after 15 seconds. This is a long time
for a single CPU unplug to occur, regardless of guest load - although
the user is *strongly* encouraged to *not* hotunplug devices from a
guest under high load - and we can be sure that something went wrong if
it takes longer than that for the guest to release the CPU (the same
can't be said about memory hotunplug - more on that in the next patch).
Timing out the unplug operation will reset the unplug state of the CPU
and allow the user to try it again, regardless of the error situation
that prevented the hotunplug to occur. Of all the not so pretty
fixes/mitigations for CPU hotunplug errors in pSeries, timing out the
operation is an admission that we have no control in the process, and
must assume the worst case if the operation doesn't succeed in a
sensible time frame.
[1] https://lists.gnu.org/archive/html/qemu-devel/2021-01/msg03353.html
[2] https://lists.gnu.org/archive/html/qemu-devel/2021-01/msg04400.html
Reported-by: Xujun Ma <xuma@redhat.com>
Fixes: https://bugzilla.redhat.com/show_bug.cgi?id=1911414
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Daniel Henrique Barboza <danielhb413@gmail.com>
Message-Id: <20210222194531.62717-5-danielhb413@gmail.com>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
The LoPAR spec provides no way for the guest kernel to report failure of
hotplug/hotunplug events. This wouldn't be bad if those operations were
granted to always succeed, but that's far for the reality.
What ends up happening is that, in the case of a failed hotunplug,
regardless of whether it was a QEMU error or a guest misbehavior, the
pSeries machine is retaining the unplug state of the device in the
running guest. This state is cleanup in machine reset, where it is
assumed that this state represents a device that is pending unplug, and
the device is hotunpluged from the board. Until the reset occurs, any
hotunplug operation of the same device is forbid because there is a
pending unplug state.
This behavior has at least one undesirable side effect. A long standing
pending unplug state is, more often than not, the result of a hotunplug
error. The user had to dealt with it, since retrying to unplug the
device is noy allowed, and then in the machine reset we're removing the
device from the guest. This means that we're failing the user twice -
failed to hotunplug when asked, then hotunplugged without notice.
Solutions to this problem range between trying to predict when the
hotunplug will fail and forbid the operation from the QEMU layer, from
opening up the IRQ queue to allow for multiple hotunplug attempts, from
telling the users to 'reboot the machine if something goes wrong'. The
first solution is flawed because we can't fully predict guest behavior
from QEMU, the second solution is a trial and error remediation that
counts on a hope that the unplug will eventually succeed, and the third
is ... well.
This patch introduces a crude, but effective solution to hotunplug
errors in the pSeries machine. For each unplug done, we'll timeout after
some time. If a certain amount of time passes, we'll cleanup the
hotunplug state from the machine. During the timeout period, any unplug
operations in the same device will still be blocked. After that, we'll
assume that the guest failed the operation, and allow the user to try
again. If the timeout is too short we'll prevent legitimate hotunplug
situations to occur, so we'll need to overestimate the regular time an
unplug operation takes to succeed to account that.
The true solution for the hotunplug errors in the pSeries machines is a
PAPR change to allow for the guest to warn the platform about it. For
now, the work done in this timeout design can be used for the new PAPR
'abort hcall' in the future, given that for both cases we'll need code
to cleanup the existing unplug states of the DRCs.
At this moment we're adding the basic wiring of the timer into the DRC.
Next patch will use the timer to timeout failed CPU hotunplugs.
Signed-off-by: Daniel Henrique Barboza <danielhb413@gmail.com>
Message-Id: <20210222194531.62717-4-danielhb413@gmail.com>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
spapr_drc_detach() is not the best name for what the function does. The
function does not detach the DRC, it makes an uncommited attempt to do
it. It'll mark the DRC as pending unplug, via the 'unplug_request'
flag, and only if the DRC state is drck->empty_state it will detach the
DRC, via spapr_drc_release().
This is a contrast with its pair spapr_drc_attach(), where the function
is indeed creating the DRC QOM object. If you know what
spapr_drc_attach() does, you can be misled into thinking that
spapr_drc_detach() is removing the DRC from QEMU internal state, which
isn't true.
The current role of this function is better described as a request for
detach, since there's no guarantee that we're going to detach the DRC in
the end. Rename the function to spapr_drc_unplug_request to reflect
what is is doing.
The initial idea was to change the name to spapr_drc_detach_request(),
and later on change the unplug_request flag to detach_request. However,
unplug_request is a migratable boolean for a long time now and renaming
it is not worth the trouble. spapr_drc_unplug_request() setting
drc->unplug_request is more natural than spapr_drc_detach_request
setting drc->unplug_request.
Reviewed-by: Greg Kurz <groug@kaod.org>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Daniel Henrique Barboza <danielhb413@gmail.com>
Message-Id: <20210222194531.62717-3-danielhb413@gmail.com>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
No need to expose the way DRCs are traversed outside of spapr_drc.c.
Signed-off-by: Greg Kurz <groug@kaod.org>
Message-Id: <20201218103400.689660-4-groug@kaod.org>
Reviewed-by: Daniel Henrique Barboza <danielhb413@gmail.com>
Tested-by: Daniel Henrique Barboza <danielhb413@gmail.com>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Documentation of object_property_iter_init() clearly stipulates that
"it is forbidden to modify the property list while iterating". But this
is exactly what we do when resetting transient DR connectors during CAS.
The call to spapr_drc_reset() can finalize the hot-unplug sequence of a
PHB or a PCI bridge, both of which will then in turn destroy their PCI
DRCs. This could potentially invalidate the iterator. It is pure luck
that this haven't caused any issues so far.
Change spapr_drc_reset() to return true if it caused a device to be
removed. Restart from scratch in this case. This can potentially
increase the overall DRC reset time, especially with a high maxmem
which generates a lot of LMB DRCs. But this kind of setup is rare,
and so is the use case of rebooting a guest while doing hot-unplug.
Signed-off-by: Greg Kurz <groug@kaod.org>
Message-Id: <20201218103400.689660-3-groug@kaod.org>
Reviewed-by: Daniel Henrique Barboza <danielhb413@gmail.com>
Tested-by: Daniel Henrique Barboza <danielhb413@gmail.com>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Non-transient DRCs are either in the empty or the ready state,
which means spapr_drc_reset() doesn't change their state. It
is thus not needed to do any checking. Call spapr_drc_reset()
unconditionally and squash spapr_drc_transient() into its
only user, spapr_drc_needed().
Signed-off-by: Greg Kurz <groug@kaod.org>
Message-Id: <20201218103400.689660-2-groug@kaod.org>
Reviewed-by: Daniel Henrique Barboza <danielhb413@gmail.com>
Tested-by: Daniel Henrique Barboza <danielhb413@gmail.com>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
All users are passing &error_abort already. Document the fact
that spapr_drc_attach() should only be passed a free DRC, which
is supposedly the case if appropriate checking is done earlier.
Signed-off-by: Greg Kurz <groug@kaod.org>
Message-Id: <20201201113728.885700-5-groug@kaod.org>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
As recommended in "qapi/error.h", return true on success and false on
failure. This allows to reduce error propagation overhead in the callers.
Signed-off-by: Greg Kurz <groug@kaod.org>
Message-Id: <20200914123505.612812-9-groug@kaod.org>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Philippe Mathieu-Daudé <philmd@redhat.com>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
spapr_drc.h includes typechecker macro boilerplate for the many different
DRC subclasses. However, most of these types don't actually have different
data in their class and/or instance, making these unneeded, unused, and in
fact a bad idea. Remove them.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
We currently don't support hotplug of devices between boot and CAS. If
this happens a CAS reboot is triggered. We detect this during CAS using
the spapr_drc_needed() function which is essentially a VMStateDescription
.needed callback. Even if the condition for CAS reboot happens to be the
same as for DRC migration, it looks wrong to piggyback a migration helper
for this.
Introduce a helper with slightly more explicit name and use it in both CAS
and DRC migration code. Since a subsequent patch will enhance this helper
to cover the case of hot unplug, let's go for spapr_drc_transient(). While
here convert spapr_hotplugged_dev_before_cas() to the "transient" wording as
well.
This doesn't change any behaviour.
Signed-off-by: Greg Kurz <groug@kaod.org>
Message-Id: <158169248180.3465937.9531405453362718771.stgit@bahia.lan>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Add support for NVDIMM devices for sPAPR. Piggyback on existing nvdimm
device interface in QEMU to support virtual NVDIMM devices for Power.
Create the required DT entries for the device (some entries have
dummy values right now).
The patch creates the required DT node and sends a hotplug
interrupt to the guest. Guest is expected to undertake the normal
DR resource add path in response and start issuing PAPR SCM hcalls.
The device support is verified based on the machine version unlike x86.
This is how it can be used ..
Ex :
For coldplug, the device to be added in qemu command line as shown below
-object memory-backend-file,id=memnvdimm0,prealloc=yes,mem-path=/tmp/nvdimm0,share=yes,size=1073872896
-device nvdimm,label-size=128k,uuid=75a3cdd7-6a2f-4791-8d15-fe0a920e8e9e,memdev=memnvdimm0,id=nvdimm0,slot=0
For hotplug, the device to be added from monitor as below
object_add memory-backend-file,id=memnvdimm0,prealloc=yes,mem-path=/tmp/nvdimm0,share=yes,size=1073872896
device_add nvdimm,label-size=128k,uuid=75a3cdd7-6a2f-4791-8d15-fe0a920e8e9e,memdev=memnvdimm0,id=nvdimm0,slot=0
Signed-off-by: Shivaprasad G Bhat <sbhat@linux.ibm.com>
Signed-off-by: Bharata B Rao <bharata@linux.ibm.com>
[Early implementation]
Message-Id: <158131058078.2897.12767731856697459923.stgit@lep8c.aus.stglabs.ibm.com>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
sysemu/sysemu.h is a rather unfocused dumping ground for stuff related
to the system-emulator. Evidence:
* It's included widely: in my "build everything" tree, changing
sysemu/sysemu.h still triggers a recompile of some 1100 out of 6600
objects (not counting tests and objects that don't depend on
qemu/osdep.h, down from 5400 due to the previous two commits).
* It pulls in more than a dozen additional headers.
Split stuff related to run state management into its own header
sysemu/runstate.h.
Touching sysemu/sysemu.h now recompiles some 850 objects. qemu/uuid.h
also drops from 1100 to 850, and qapi/qapi-types-run-state.h from 4400
to 4200. Touching new sysemu/runstate.h recompiles some 500 objects.
Since I'm touching MAINTAINERS to add sysemu/runstate.h anyway, also
add qemu/main-loop.h.
Suggested-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Message-Id: <20190812052359.30071-30-armbru@redhat.com>
Reviewed-by: Alex Bennée <alex.bennee@linaro.org>
[Unbreak OS-X build]
In my "build everything" tree, changing hw/qdev-properties.h triggers
a recompile of some 2700 out of 6600 objects (not counting tests and
objects that don't depend on qemu/osdep.h).
Many places including hw/qdev-properties.h (directly or via hw/qdev.h)
actually need only hw/qdev-core.h. Include hw/qdev-core.h there
instead.
hw/qdev.h is actually pointless: all it does is include hw/qdev-core.h
and hw/qdev-properties.h, which in turn includes hw/qdev-core.h.
Replace the remaining uses of hw/qdev.h by hw/qdev-properties.h.
While there, delete a few superfluous inclusions of hw/qdev-core.h.
Touching hw/qdev-properties.h now recompiles some 1200 objects.
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: "Daniel P. Berrangé" <berrange@redhat.com>
Cc: Eduardo Habkost <ehabkost@redhat.com>
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eduardo Habkost <ehabkost@redhat.com>
Message-Id: <20190812052359.30071-22-armbru@redhat.com>
Some of the generated qapi-types-MODULE.h are included all over the
place. Changing a QAPI type can trigger massive recompiling. Top
scorers recompile more than 1000 out of some 6600 objects (not
counting tests and objects that don't depend on qemu/osdep.h):
6300 qapi/qapi-builtin-types.h
5700 qapi/qapi-types-run-state.h
3900 qapi/qapi-types-common.h
3300 qapi/qapi-types-sockets.h
3000 qapi/qapi-types-misc.h
3000 qapi/qapi-types-crypto.h
3000 qapi/qapi-types-job.h
3000 qapi/qapi-types-block-core.h
2800 qapi/qapi-types-block.h
1300 qapi/qapi-types-net.h
Clean up headers to include generated QAPI headers only where needed.
Impact is negligible except for hw/qdev-properties.h.
This header includes qapi/qapi-types-block.h and
qapi/qapi-types-misc.h. They are used only in expansions of property
definition macros such as DEFINE_PROP_BLOCKDEV_ON_ERROR() and
DEFINE_PROP_OFF_AUTO(). Moving their inclusion from
hw/qdev-properties.h to the users of these macros avoids pointless
recompiles. This is how other property definition macros, such as
DEFINE_PROP_NETDEV(), already work.
Improves things for some of the top scorers:
3600 qapi/qapi-types-common.h
2800 qapi/qapi-types-sockets.h
900 qapi/qapi-types-misc.h
2200 qapi/qapi-types-crypto.h
2100 qapi/qapi-types-job.h
2100 qapi/qapi-types-block-core.h
270 qapi/qapi-types-block.h
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Philippe Mathieu-Daudé <philmd@redhat.com>
Tested-by: Philippe Mathieu-Daudé <philmd@redhat.com>
Message-Id: <20190812052359.30071-3-armbru@redhat.com>
This makes some minor cleanups to spapr_drc_populate_dt(), renaming it to
the shorter and more idiomatic spapr_dt_drc() along the way.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
Acked-by: Michael S. Tsirkin <mst@redhat.com>
The qemu coding standard is to use CamelCase for type and structure names,
and the pseries code follows that... sort of. There are quite a lot of
places where we bend the rules in order to preserve the capitalization of
internal acronyms like "PHB", "TCE", "DIMM" and most commonly "sPAPR".
That was a bad idea - it frequently leads to names ending up with hard to
read clusters of capital letters, and means they don't catch the eye as
type identifiers, which is kind of the point of the CamelCase convention in
the first place.
In short, keeping type identifiers look like CamelCase is more important
than preserving standard capitalization of internal "words". So, this
patch renames a heap of spapr internal type names to a more standard
CamelCase.
In addition to case changes, we also make some other identifier renames:
VIOsPAPR* -> SpaprVio*
The reverse word ordering was only ever used to mitigate the capital
cluster, so revert to the natural ordering.
VIOsPAPRVTYDevice -> SpaprVioVty
VIOsPAPRVLANDevice -> SpaprVioVlan
Brevity, since the "Device" didn't add useful information
sPAPRDRConnector -> SpaprDrc
sPAPRDRConnectorClass -> SpaprDrcClass
Brevity, and makes it clearer this is the same thing as a "DRC"
mentioned in many other places in the code
This is 100% a mechanical search-and-replace patch. It will, however,
conflict with essentially any and all outstanding patches touching the
spapr code.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Michael Roth <mdroth@linux.vnet.ibm.com>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Greg Kurz <groug@kaod.org>
Message-Id: <155059670389.1466090.10015601248906623076.stgit@bahia.lab.toulouse-stg.fr.ibm.com>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
All DRC subtypes have been converted to generate the FDT fragment at
configure connector time instead of attach time. The fdt and fdt_offset
arguments of spapr_drc_attach() aren't needed anymore. Drop them and
make the implementation of the dt_populate() method mandatory.
Signed-off-by: Greg Kurz <groug@kaod.org>
Message-Id: <155059667853.1466090.16527852453054217565.stgit@bahia.lab.toulouse-stg.fr.ibm.com>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
The current logic is to provide the FDT fragment when attaching a device
to a DRC. This works perfectly fine for our current hotplug support, but
soon we will add support for PHB hotplug which has some constraints, that
CPU, PCI and LMB devices don't seem to have.
The first constraint is that the "ibm,dma-window" property of the PHB
node requires the IOMMU to be configured, ie, spapr_tce_table_enable()
has been called, which happens during PHB reset. It is okay in the case
of hotplug since the device is reset before the hotplug handler is
called. On the contrary with coldplug, the hotplug handler is called
first and device is only reset during the initial system reset. Trying
to create the FDT fragment on the hotplug path in this case, would
result in somthing like this:
ibm,dma-window = < 0x80000000 0x00 0x00 0x00 0x00 >;
This will cause linux in the guest to panic, by simply removing and
re-adding the PHB using the drmgr command:
page = alloc_pages_node(nid, GFP_KERNEL, get_order(sz));
if (!page)
panic("iommu_init_table: Can't allocate %ld bytes\n", sz);
The second and maybe more problematic constraint is that the
"interrupt-map" property needs to reference the interrupt controller
node using the very same phandle that SLOF has already exposed to the
guest. QEMU requires SLOF to call the private KVMPPC_H_UPDATE_DT hcall
at some point to know about this phandle. With the latest QEMU and SLOF,
this happens when SLOF gets quiesced. This means that if the PHB gets
hotplugged after CAS but before SLOF quiesce, then we're sure that the
phandle is not known when the hotplug handler is called.
The FDT is only needed when the guest first invokes RTAS to configure
the connector actually, long after SLOF quiesce. Let's postpone the
creation of FDT fragments for PHBs to rtas_ibm_configure_connector().
Since we only need this for PHBs, introduce a new method in the base
DRC class for that. DRC subtypes will be converted to use it in
subsequent patches.
Allow spapr_drc_attach() to be passed a NULL fdt argument if the method
is available. When all DRC subtypes have been converted, the fdt argument
will eventually disappear.
Signed-off-by: Greg Kurz <groug@kaod.org>
Message-Id: <155059665823.1466090.18358845122627355537.stgit@bahia.lab.toulouse-stg.fr.ibm.com>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
In my "build everything" tree, a change to the types in
qapi-schema.json triggers a recompile of about 4800 out of 5100
objects.
The previous commit split up qmp-commands.h, qmp-event.h, qmp-visit.h,
qapi-types.h. Each of these headers still includes all its shards.
Reduce compile time by including just the shards we actually need.
To illustrate the benefits: adding a type to qapi/migration.json now
recompiles some 2300 instead of 4800 objects. The next commit will
improve it further.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Message-Id: <20180211093607.27351-24-armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Marc-André Lureau <marcandre.lureau@redhat.com>
[eblake: rebase to master]
Signed-off-by: Eric Blake <eblake@redhat.com>
This patch is a follow up on the discussions made in patch
"hw/ppc: disable hotplug before CAS is completed" that can be
found at [1].
At this moment, we do not support CPU/memory hotplug in early
boot stages, before CAS. When a hotplug occurs, the event is logged
in an internal RTAS event log queue and an IRQ pulse is fired. In
regular conditions, the guest handles the interrupt by executing
check_exception, fetching the generated hotplug event and enabling
the device for use.
In early boot, this IRQ isn't caught (SLOF does not handle hotplug
events), leaving the event in the rtas event log queue. If the guest
executes check_exception due to another hotplug event, the re-assertion
of the IRQ ends up de-queuing the first hotplug event as well. In short,
a device hotplugged before CAS is considered coldplugged by SLOF.
This leads to device misbehavior and, in some cases, guest kernel
Ooops when trying to unplug the device.
A proper fix would be to turn every device hotplugged before CAS
as a colplugged device. This is not trivial to do with the current
code base though - the FDT is written in the guest memory at
ppc_spapr_reset and can't be retrieved without adding extra state
(fdt_size for example) that will need to managed and migrated. Adding
the hotplugged DT in the middle of CAS negotiation via the updated DT
tree works with CPU devs, but panics the guest kernel at boot. Additional
analysis would be necessary for LMBs and PCI devices. There are
questions to be made in QEMU/SLOF/kernel level about how we can make
this change in a sustainable way.
With Linux guests, a fix would be the kernel executing check_exception
at boot time, de-queueing the events that happened in early boot and
processing them. However, even if/when the newer kernels start
fetching these events at boot time, we need to take care of older
kernels that won't be doing that.
This patch works around the situation by issuing a CAS reset if a hotplugged
device is detected during CAS:
- the DRC conditions that warrant a CAS reset is the same as those that
triggers a DRC migration - the DRC must have a device attached and
the DRC state is not equal to its ready_state. With that in mind, this
patch makes use of 'spapr_drc_needed' to determine if a CAS reset
is needed.
- In the middle of CAS negotiations, the function
'spapr_hotplugged_dev_before_cas' goes through all the DRCs to see
if there are any DRC that requires a reset, using spapr_drc_needed. If
that happens, returns '1' in 'spapr_h_cas_compose_response' which will set
spapr->cas_reboot to true, causing the machine to reboot.
No changes are made for coldplug devices.
[1] http://lists.nongnu.org/archive/html/qemu-devel/2017-08/msg02855.html
Signed-off-by: Daniel Henrique Barboza <danielhb@linux.vnet.ibm.com>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
According to PAPR, the DR-indicator should only be valid for physical DRCs,
not logical DRCs. At the moment we implement it for all DRCs, so restrict
it to physical ones only.
We move the state to the physical DRC subclass, which means adding some
QOM boilerplate to handle the newly distinct type.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Daniel Barboza <danielhb@linux.vnet.ibm.com>
Tested-by: Daniel Barboza <danielhb@linux.vnet.ibm.com>
Most of the time, the state of a DRC object is contained in the single
'state' variable. However, during the transition from UNISOLATE to
CONFIGURED state requires multiple calls to the ibm,configure-connector
RTAS call to retrieve the device tree for the attached device. We need
some extra state to keep track of where we're up to in delivering the
device tree information to the guest.
Currently that extra state is in a sPAPRConfigureConnectorState
substructure which is only allocated when we're in the middle of the
configure connector process. That sounds like a good idea, but the extra
state is only two integers - on many platforms that will take up the same
room as the (maybe NULL) ccs pointer even before malloc() overhead. Plus
it's another object whose lifetime we need to manage. In short, it's not
worth it.
So, fold the sPAPRConfigureConnectorState substructure directly into the
DRC object.
Previously the structure was allocated lazily when the configure-connector
call discovers it's not there. Now, we need to initialize the subfields
pre-emptively, as soon as we enter UNISOLATE state.
Although it's not strictly necessary (the field values should only ever
be consulted when in UNISOLATE state), we try to keep them at -1 when in
other states, as a debugging aid.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Daniel Barboza <danielhb@linux.vnet.ibm.com>
Tested-by: Daniel Barboza <danielhb@linux.vnet.ibm.com>
Each DRC has three fields describing its state: isolation_state,
allocation_state and configured. At first this seems like a reasonable
representation, since its based directly on the PAPR defined
isolation-state and allocation-state indicators. However:
* Only a few combinations of the two fields' values are permitted
* allocation_state isn't used at all for physical DRCs
* The indicators are write only so they don't really have a well
defined current value independent of each other
This replaces these variables with a single state variable, whose names
and numbers are based on the diagram in LoPAPR section 13.4. Along with
this we add code to check the current state on various operations and make
sure the requested transition is permitted.
Strictly speaking, this makes guest visible changes to behaviour (since we
probably allowed some transitions we shouldn't have before). However, a
hypothetical guest broken by that wasn't PAPR compliant, and probably
wouldn't have worked under PowerVM.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Daniel Barboza <danielhb@linux.vnet.ibm.com>
Tested-by: Daniel Barboza <danielhb@linux.vnet.ibm.com>
'awaiting_release' indicates that the host has requested an unplug of the
device attached to the DRC, but the guest has not (yet) put the device
into a state where it is safe to complete removal.
1. Rename it to 'unplug_requested' which to me at least is clearer
2. Remove the ->release_pending() method used to check this from outside
spapr_drc.c. The method only plausibly has one implementation, so use
a plain function (spapr_drc_unplug_requested()) instead.
3. Remove it from the migration stream. Attempting to migrate mid-unplug
is broken not just for spapr - in general management has no good way to
determine if the device should be present on the destination or not. So,
until that's fixed, there's no point adding extra things to the stream.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
Tested-by: Daniel Barboza <danielhb@linux.vnet.ibm.com>
This function has two unused parameters - remove them.
It also sets awaiting_release on all paths, except one. On that path
setting it is harmless, since it will be immediately cleared by
spapr_drc_release(). So factor it out of the if statements.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
Tested-by: Daniel Barboza <danielhb@linux.vnet.ibm.com>
The awaiting_allocation flag in the DRC was introduced by aab9913
"spapr_drc: Prevent detach racing against attach for CPU DR", allegedly to
prevent a guest crash on racing attach and detach. Except.. information
from the BZ actually suggests a qemu crash, not a guest crash. And there
shouldn't be a problem here anyway: if the guest has already moved the DRC
away from UNUSABLE state, the detach would already be deferred, and if it
hadn't it should be safe to detach it (the guest should fail gracefully
when it attempts to change the allocation state).
I think this was probably just a bandaid for some other problem in the
state management. So, remove awaiting_allocation and associated code.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Laurent Vivier <lvivier@redhat.com>
Reviewed-by: Greg Kurz <groug@kaod.org>
Tested-by: Greg Kurz <groug@kaod.org>
Tested-by: Daniel Barboza <danielhb@linux.vnet.ibm.com>
When migrating a guest which has already had devices hotplugged,
libvirt typically starts the destination qemu with -incoming defer,
adds those hotplugged devices with qmp, then initiates the incoming
migration.
This causes problems for the management of spapr DRC state. Because
the device is treated as hotplugged, it goes into a DRC state for a
device immediately after it's plugged, but before the guest has
acknowledged its presence. However, chances are the guest on the
source machine *has* acknowledged the device's presence and configured
it.
If the source has fully configured the device, then DRC state won't be
sent in the migration stream: for maximum migration compatibility with
earlier versions we don't migrate DRCs in coldplug-equivalent state.
That means that the DRC effectively changes state over the migrate,
causing problems later on.
In addition, logging hotplug events for these devices isn't what we
want because a) those events should already have been issued on the
source host and b) the event queue should get wiped out by the
incoming state anyway.
In short, what we really want is to treat devices added before an
incoming migration as if they were coldplugged.
To do this, we first add a spapr_drc_hotplugged() helper which
determines if the device is hotplugged in the sense relevant for DRC
state management. We only send hotplug events when this is true.
Second, when we add a device which isn't hotplugged in this sense, we
force a reset of the DRC state - this ensures the DRC is in a
coldplug-equivalent state (there isn't usually a system reset between
these device adds and the incoming migration).
This is based on an earlier patch by Laurent Vivier, cleaned up and
extended.
Signed-off-by: Laurent Vivier <lvivier@redhat.com>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
Tested-by: Daniel Barboza <danielhb@linux.vnet.ibm.com>
spapr_drc_attach() has a 'coldplug' parameter which sets the DRC into
configured state initially, instead of the usual ISOLATED/UNUSABLE state.
It turns out this is unnecessary: although coldplugged devices do need to
be in CONFIGURED state once the guest starts, that will already be
accomplished by the reset code which will move DRCs for already plugged
devices into a coldplug equivalent state.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Laurent Vivier <lvivier@redhat.com>
Reviewed-by: Greg Kurz <groug@kaod.org>
At the moment, spapr_drc_release() has an ugly switch on the DRC type to
call the right, device-specific release function. This cleans it up by
doing that via a proper QOM method.
It's still arguably an abstraction violation for the DRC code to call into
the specific device code, but one mess at a time.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Laurent Vivier <lvivier@redhat.com>
Reviewed-by: Greg Kurz <groug@kaod.org>
There are substantial differences in the various paths through
set_isolation_state(), both for setting to ISOLATED versus UNISOLATED
state and for logical versus physical DRCs.
So, split the set_isolation_state() method into isolate() and unisolate()
methods, and give it different implementations for the two DRC types.
Factor some minimal common checks, including for valid indicator values
(which we weren't previously checking) into rtas_set_isolation_state().
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Michael Roth <mdroth@linux.vnet.ibm.com>
The allocation-state indicator should only actually be implemented for
"logical" DRCs, not physical ones. Factor a check for this, and also for
valid indicator state values into rtas_set_allocation_state(). Because
they don't exist for physical DRCs, there's no reason that we'd ever want
more than one method implementation, so it can just be a plain function.
In addition, the setting to USABLE and setting to UNUSABLE paths in
set_allocation_state() don't actually have much in common. So, split the
method separate functions for each parameter value (drc_set_usable()
and drc_set_unusable()).
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Michael Roth <mdroth@linux.vnet.ibm.com>
The 'signalled' field in the DRC appears to be entirely a torturous
workaround for the fact that PCI devices were started in UNISOLATED state
for unclear reasons.
1) 'signalled' is already meaningless for logical (so far, all non PCI)
DRCs. It's always set to true (at least at any point it might be tested),
and can't be assigned any real meaning due to the way signalling works for
logical DRCs.
2) For PCI DRCs, the only time signalled would be false is when non-zero
functions of a multifunction device are hotplugged, followed by function
zero (the other way around is explicitly not permitted). In that case the
secondary function DRCs are attached, but the notification isn't sent to
the guest until function 0 is plugged.
3) signalled being false is used to allow a DRC detach to switch mode
back to ISOLATED state, which allows a secondary function to be hotplugged
then unplugged with function 0 never inserted. Without this a secondary
function starting in UNISOLATED state couldn't be detached again without
function 0 being inserted, all the functions configured by the guest, then
sent back to ISOLATED state.
4) But now that PCI DRCs start in ISOLATED state, there's nothing to be
done. If the guest doesn't get the notification, it won't switch the
device to UNISOLATED state, so nothing prevents it from being unplugged.
If the guest does move it to UNISOLATED state without the signal (due to
a manual drmgr call, for instance) then it really isn't safe to unplug it.
So, this patch removes the signalled variable and all code related to it.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Michael Roth <mdroth@linux.vnet.ibm.com>
This reverts commit fe6824d126.
Conflicts hw/ppc/spapr_drc.c, because get_index() has been renamed
spapr_get_index().
This didn't fix the problem. Once the hotplug has been started
some memory is allocated and some structures are allocated.
We don't free it when we ignore the unplug, and we can't because
they can be in use by the kernel.
Signed-off-by: Laurent Vivier <lvivier@redhat.com>
Tested-by: Daniel Barboza <danielhb@linux.vnet.ibm.com>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
DRC objects have a get_name method which returns the DRC name generated
when the DRC is created. Replace that with a fixed spapr_drc_name()
function which generates the name on the fly from other information. This
means:
* We get rid of a method with only one implementation, and only local
callers
* We don't have to carry the name string around for the lifetime of the
DRC
* We use information added to the class structure to generate the name
in standard format, so we don't need an explicit switch on drc type
any more
We also eliminate the 'name' property; it's basically useless since the
only information in it can easily be deduced from other things.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Michael Roth <mdroth@linux.vnet.ibm.com>
Acked-by: Michael Roth <mdroth@linux.vnet.ibm.com>
DRC objects have attach & detach methods, but there's only one
implementation. Although there are some differences in its behaviour for
different DRC types, the overall structure is the same, so while we might
want different method implementations for some parts, we're unlikely to
want them for the top-level functions.
So, replace them with direct function calls.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Michael Roth <mdroth@linux.vnet.ibm.com>
Acked-by: Michael Roth <mdroth@linux.vnet.ibm.com>
There are 3 types of "indicator" associated with hotplug in the PAPR spec
the "allocation state", "isolation state" and "DR-indicator". The first
two are intimately tied to the various state transitions associated with
hotplug. The DR-indicator, however, is different and simpler.
It's basically just a guest controlled variable which can be used by the
guest to flag state or problems associated with a device. The idea is that
the hypervisor can use it to present information back on management
consoles (on some machines with PowerVM it may even control physical LEDs
on the machine case associated with the relevant device).
For that reason, there's only ever likely to be a single update
implementation so the set_indicator_state method isn't useful. Replace it
with a direct function call.
While we're there, make some small associated cleanups:
* PAPR doesn't use the term "indicator state", just "DR-indicator" and
the allocation state and isolation state are also considered "indicators".
Rename things to be less confusing
* Fold set_indicator_state() and rtas_set_indicator_state() into a single
rtas_set_dr_indicator() function.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Michael Roth <mdroth@linux.vnet.ibm.com>
Acked-by: Michael Roth <mdroth@linux.vnet.ibm.com>
DRC classes have an entity_sense method to determine (in a specific PAPR
sense) the presence or absence of a device plugged into a DRC. However,
we only have one implementation of the method, which explicitly tests for
different DRC types. This changes it to instead have different method
implementations for the two cases: "logical" and "physical" DRCs.
While we're at it, the entity sense method always returns RTAS_OUT_SUCCESS,
and the interesting value is returned via pass-by-reference. Simplify this
to directly return the value we care about
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Michael Roth <mdroth@linux.vnet.ibm.com>
Acked-by: Michael Roth <mdroth@linux.vnet.ibm.com>
This function was used in generating the device tree. However, now that
we have different QOM types for different DRC types we can easily store
the information we need in the class structure and avoid this specialized
lookup function.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Michael Roth <mdroth@linux.vnet.ibm.com>
Acked-by: Michael Roth <mdroth@linux.vnet.ibm.com>
Currently the sPAPRMachineState contains a list of sPAPRConfigureConnector
structures which store intermediate state for the ibm,configure-connector
RTAS call.
This was an attempt to separate this state from the core of the DRC state.
However the configure connector process is intimately tied to the DRC
model, so there's really no point trying to have two levels of interface
here.
Moving the configure-connector state into its corresponding DRC allows
removal of a number of helpers for maintaining the anciliary list.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Michael Roth <mdroth@linux.vnet.ibm.com>
Acked-by: Michael Roth <mdroth@linux.vnet.ibm.com>
* Change names to something less ludicrously verbose
* Now that we have QOM subclasses for the different DRC types, use a QOM
typename instead of a PAPR type value parameter
The latter allows removal of the get_type_shift() helper.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Michael Roth <mdroth@linux.vnet.ibm.com>
Acked-by: Michael Roth <mdroth@linux.vnet.ibm.com>
Currently we only have a single QOM type for all DRCs, but lots of
places where we switch behaviour based on the DRC's PAPR defined type.
This is a poor use of our existing type system.
So, instead create QOM subclasses for each PAPR defined DRC type. We
also introduce intermediate subclasses for physical and logical DRCs,
a division which will be useful later on.
Instead of being stored in the DRC object itself, the PAPR type is now
stored in the class structure. There are still many places where we
switch directly on the PAPR type value, but this at least provides the
basis to start to remove those.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Michael Roth <mdroth@linux.vnet.ibm.com>
Acked-by: Michael Roth <mdroth@linux.vnet.ibm.com>
These two methods only have one implementation, and the spec they're
implementing means any other implementation is unlikely, verging on
impossible.
So replace them with simple functions.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Laurent Vivier <lvivier@redhat.com>
Tested-by: Daniel Barboza <danielhb@linux.vnet.ibm.com>
DRConnectorClass has a set_configured method, however:
* There is only one implementation, and only ever likely to be one
* There's exactly one caller, and that's (now) local
* The implementation is very straightforward
So abolish the method entirely, and just open-code what we need.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Laurent Vivier <lvivier@redhat.com>
Reviewed-by: Greg Kurz <groug@kaod.org>
Tested-by: Daniel Barboza <danielhb@linux.vnet.ibm.com>
The DRConnectorClass includes a get_fdt method. However
* There's only one implementation, and there's only likely to ever be one
* Both callers are local to spapr_drc
* Each caller only uses one half of the actual implementation
So abolish get_fdt() entirely, and just open-code what we need.
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Reviewed-by: Laurent Vivier <lvivier@redhat.com>
Reviewed-by: Greg Kurz <groug@kaod.org>
Tested-by: Daniel Barboza <danielhb@linux.vnet.ibm.com>
The pointer drc->detach_cb is being used as a way of informing
the detach() function inside spapr_drc.c which cb to execute. This
information can also be retrieved simply by checking drc->type and
choosing the right callback based on it. In this context, detach_cb
is redundant information that must be managed.
After the previous spapr_lmb_release change, no detach_cb_opaques
are being used by any of the three callbacks functions. This is
yet another information that is now unused and, on top of that, can't
be migrated either.
This patch makes the following changes:
- removal of detach_cb_opaque. the 'opaque' argument was removed from
the callbacks and from the detach() function of sPAPRConnectorClass. The
attribute detach_cb_opaque of sPAPRConnector was removed.
- removal of detach_cb from the detach() call. The function pointer
detach_cb of sPAPRConnector was removed. detach() now uses a
switch(drc->type) to execute the apropriate callback. To achieve this,
spapr_core_release, spapr_lmb_release and spapr_phb_remove_pci_device_cb
callbacks were made public to be visible inside detach().
Signed-off-by: Daniel Henrique Barboza <danielhb@linux.vnet.ibm.com>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
If, once the kernel has booted, we try to remove a memory
hotplugged while the kernel was not started, QEMU crashes on
an assert:
qemu-system-ppc64: hw/virtio/vhost.c:651:
vhost_commit: Assertion `r >= 0' failed.
...
#4 in vhost_commit
#5 in memory_region_transaction_commit
#6 in pc_dimm_memory_unplug
#7 in spapr_memory_unplug
#8 spapr_machine_device_unplug
#9 in hotplug_handler_unplug
#10 in spapr_lmb_release
#11 in detach
#12 in set_allocation_state
#13 in rtas_set_indicator
...
If we take a closer look to the guest kernel log, we can see when
we try to unplug the memory:
pseries-hotplug-mem: Attempting to hot-add 4 LMB(s)
What happens:
1- The kernel has ignored the memory hotplug event because
it was not started when it was generated.
2- When we hot-unplug the memory,
QEMU starts to remove the memory,
generates an hot-unplug event,
and signals the kernel of the incoming new event
3- as the kernel is started, on the QEMU signal, it reads
the event list, decodes the hotplug event and tries to
finish the hotplugging.
4- QEMU receive the the hotplug notification while it
is trying to hot-unplug the memory. This moves the memory
DRC to an invalid state
This patch prevents this by not allowing to set the allocation
state to USABLE while the DRC is awaiting release.
RHBZ: https://bugzilla.redhat.com/show_bug.cgi?id=1432382
Signed-off-by: Laurent Vivier <lvivier@redhat.com>
Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
Tracked down with an ugly, brittle and probably buggy Perl script.
Also move includes converted to <...> up so they get included before
ours where that's obviously okay.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Tested-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Richard Henderson <rth@twiddle.net>
