When running with "dynamic-memslots=off", we enter
virtio_mem_activate_memslots_to_plug() to return immediately again
because "vmem->dynamic_memslots == false". However, the compiler might
not optimize out calculating start_idx+end_idx, where we divide by
vmem->memslot_size. In such a configuration, the memslot size is 0 and
we'll get a division by zero:
(qemu) qom-set vmem0 requested-size 3G
(qemu) q35.sh: line 38: 622940 Floating point exception(core dumped)
The same is true for virtio_mem_deactivate_unplugged_memslots(), however
we never really reach that code without a prior
virtio_mem_activate_memslots_to_plug() call.
Let's fix it by simply calling these functions only with
"dynamic-memslots=on".
This was found when using a debug build of QEMU.
Message-ID: <20231023111341.219317-1-david@redhat.com>
Reprted-by: Mario Casquero <mcasquer@redhat.com>
Fixes: 177f9b1ee4 ("virtio-mem: Expose device memory dynamically via multiple memslots if enabled")
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Reviewed-by: Maciej S. Szmigiero <maciej.szmigiero@oracle.com>
Tested-by: Mario Casquero <mcasquer@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
There is no strong requirement that the size has to be multiples of the
requested alignment, let's drop it. This is a preparation for hv-baloon.
Signed-off-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Maciej S. Szmigiero <maciej.szmigiero@oracle.com>
Let's support empty memory devices -- memory devices that don't have a
memory device region in the current configuration. hv-balloon with an
optional memdev is the primary use case.
Signed-off-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Maciej S. Szmigiero <maciej.szmigiero@oracle.com>
Let's mark the memslot alias memory regions as unmergable, such that
flatview and vhost won't merge adjacent memory region aliases and we can
atomically map/unmap individual aliases without affecting adjacent
alias memory regions.
This handles vhost and vfio in multiple-memslot mode correctly (which do
not support atomic memslot updates) and avoids the temporary removal of
large memslots, which can be an expensive operation. For example, vfio
might have to unpin + repin a lot of memory, which is undesired.
Message-ID: <20230926185738.277351-19-david@redhat.com>
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Let's allow for marking memory regions unmergeable, to teach
flatview code and vhost to not merge adjacent aliases to the same memory
region into a larger memory section; instead, we want separate aliases to
stay separate such that we can atomically map/unmap aliases without
affecting other aliases.
This is desired for virtio-mem mapping device memory located on a RAM
memory region via multiple aliases into a memory region container,
resulting in separate memslots that can get (un)mapped atomically.
As an example with virtio-mem, the layout would look something like this:
[...]
0000000240000000-00000020bfffffff (prio 0, i/o): device-memory
0000000240000000-000000043fffffff (prio 0, i/o): virtio-mem
0000000240000000-000000027fffffff (prio 0, ram): alias memslot-0 @mem2 0000000000000000-000000003fffffff
0000000280000000-00000002bfffffff (prio 0, ram): alias memslot-1 @mem2 0000000040000000-000000007fffffff
00000002c0000000-00000002ffffffff (prio 0, ram): alias memslot-2 @mem2 0000000080000000-00000000bfffffff
[...]
Without unmergable memory regions, all three memslots would get merged into
a single memory section. For example, when mapping another alias (e.g.,
virtio-mem-memslot-3) or when unmapping any of the mapped aliases,
memory listeners will first get notified about the removal of the big
memory section to then get notified about re-adding of the new
(differently merged) memory section(s).
In an ideal world, memory listeners would be able to deal with that
atomically, like KVM nowadays does. However, (a) supporting this for other
memory listeners (vhost-user, vfio) is fairly hard: temporary removal
can result in all kinds of issues on concurrent access to guest memory;
and (b) this handling is undesired, because temporarily removing+readding
can consume quite some time on bigger memslots and is not efficient
(e.g., vfio unpinning and repinning pages ...).
Let's allow for marking a memory region unmergeable, such that we
can atomically (un)map aliases to the same memory region, similar to
(un)mapping individual DIMMs.
Similarly, teach vhost code to not redo what flatview core stopped doing:
don't merge such sections. Merging in vhost code is really only relevant
for handling random holes in boot memory where; without this merging,
the vhost-user backend wouldn't be able to mmap() some boot memory
backed on hugetlb.
We'll use this for virtio-mem next.
Message-ID: <20230926185738.277351-18-david@redhat.com>
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Having large virtio-mem devices that only expose little memory to a VM
is currently a problem: we map the whole sparse memory region into the
guest using a single memslot, resulting in one gigantic memslot in KVM.
KVM allocates metadata for the whole memslot, which can result in quite
some memory waste.
Assuming we have a 1 TiB virtio-mem device and only expose little (e.g.,
1 GiB) memory, we would create a single 1 TiB memslot and KVM has to
allocate metadata for that 1 TiB memslot: on x86, this implies allocating
a significant amount of memory for metadata:
(1) RMAP: 8 bytes per 4 KiB, 8 bytes per 2 MiB, 8 bytes per 1 GiB
-> For 1 TiB: 2147483648 + 4194304 + 8192 = ~ 2 GiB (0.2 %)
With the TDP MMU (cat /sys/module/kvm/parameters/tdp_mmu) this gets
allocated lazily when required for nested VMs
(2) gfn_track: 2 bytes per 4 KiB
-> For 1 TiB: 536870912 = ~512 MiB (0.05 %)
(3) lpage_info: 4 bytes per 2 MiB, 4 bytes per 1 GiB
-> For 1 TiB: 2097152 + 4096 = ~2 MiB (0.0002 %)
(4) 2x dirty bitmaps for tracking: 2x 1 bit per 4 KiB page
-> For 1 TiB: 536870912 = 64 MiB (0.006 %)
So we primarily care about (1) and (2). The bad thing is, that the
memory consumption *doubles* once SMM is enabled, because we create the
memslot once for !SMM and once for SMM.
Having a 1 TiB memslot without the TDP MMU consumes around:
* With SMM: 5 GiB
* Without SMM: 2.5 GiB
Having a 1 TiB memslot with the TDP MMU consumes around:
* With SMM: 1 GiB
* Without SMM: 512 MiB
... and that's really something we want to optimize, to be able to just
start a VM with small boot memory (e.g., 4 GiB) and a virtio-mem device
that can grow very large (e.g., 1 TiB).
Consequently, using multiple memslots and only mapping the memslots we
really need can significantly reduce memory waste and speed up
memslot-related operations. Let's expose the sparse RAM memory region using
multiple memslots, mapping only the memslots we currently need into our
device memory region container.
The feature can be enabled using "dynamic-memslots=on" and requires
"unplugged-inaccessible=on", which is nowadays the default.
Once enabled, we'll auto-detect the number of memslots to use based on the
memslot limit provided by the core. We'll use at most 1 memslot per
gigabyte. Note that our global limit of memslots accross all memory devices
is currently set to 256: even with multiple large virtio-mem devices,
we'd still have a sane limit on the number of memslots used.
The default is to not dynamically map memslot for now
("dynamic-memslots=off"). The optimization must be enabled manually,
because some vhost setups (e.g., hotplug of vhost-user devices) might be
problematic until we support more memslots especially in vhost-user backends.
Note that "dynamic-memslots=on" is just a hint that multiple memslots
*may* be used for internal optimizations, not that multiple memslots
*must* be used. The actual number of memslots that are used is an
internal detail: for example, once memslot metadata is no longer an
issue, we could simply stop optimizing for that. Migration source and
destination can differ on the setting of "dynamic-memslots".
Message-ID: <20230926185738.277351-17-david@redhat.com>
Reviewed-by: Maciej S. Szmigiero <maciej.szmigiero@oracle.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
It's cleaner and future-proof to just have other state that depends on the
bitmap state to be updated as soon as possible when restoring the bitmap.
So factor out informing RamDiscardListener into a functon and call it in
case of early migration right after we restored the bitmap.
Message-ID: <20230926185738.277351-16-david@redhat.com>
Reviewed-by: Maciej S. Szmigiero <maciej.szmigiero@oracle.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
We really only care about the RAM memory region not being mapped into
an address space yet as long as we're still setting up the
RamDiscardManager. Once mapped into an address space, memory notifiers
would get notified about such a region and any attempts to modify the
RamDiscardManager would be wrong.
While "mapped into an address space" is easy to check for RAM regions that
are mapped directly (following the ->container links), it's harder to
check when such regions are mapped indirectly via aliases. For now, we can
only detect that a region is mapped through an alias (->mapped_via_alias),
but we don't have a handle on these aliases to follow all their ->container
links to test if they are eventually mapped into an address space.
So relax the assertion in memory_region_set_ram_discard_manager(),
remove the check in memory_region_get_ram_discard_manager() and clarify
the doc.
Message-ID: <20230926185738.277351-14-david@redhat.com>
Reviewed-by: Maciej S. Szmigiero <maciej.szmigiero@oracle.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
We want to support memory devices that can automatically decide how many
memslots they will use. In the worst case, they have to use a single
memslot.
The target use cases are virtio-mem and the hyper-v balloon.
Let's calculate a reasonable limit such a memory device may use, and
instruct the device to make a decision based on that limit. Use a simple
heuristic that considers:
* A memslot soft-limit for all memory devices of 256; also, to not
consume too many memslots -- which could harm performance.
* Actually still free and unreserved memslots
* The percentage of the remaining device memory region that memory device
will occupy.
Further, while we properly check before plugging a memory device whether
there still is are free memslots, we have other memslot consumers (such as
boot memory, PCI BARs) that don't perform any checks and might dynamically
consume memslots without any prior reservation. So we might succeed in
plugging a memory device, but once we dynamically map a PCI BAR we would
be in trouble. Doing accounting / reservation / checks for all such
users is problematic (e.g., sometimes we might temporarily split boot
memory into two memslots, triggered by the BIOS).
We use the historic magic memslot number of 509 as orientation to when
supporting 256 memory devices -> memslots (leaving 253 for boot memory and
other devices) has been proven to work reliable. We'll fallback to
suggesting a single memslot if we don't have at least 509 total memslots.
Plugging vhost devices with less than 509 memslots available while we
have memory devices plugged that consume multiple memslots due to
automatic decisions can be problematic. Most configurations might just fail
due to "limit < used + reserved", however, it can also happen that these
memory devices would suddenly consume memslots that would actually be
required by other memslot consumers (boot, PCI BARs) later. Note that this
has always been sketchy with vhost devices that support only a small number
of memslots; but we don't want to make it any worse.So let's keep it simple
and simply reject plugging such vhost devices in such a configuration.
Eventually, all vhost devices that want to be fully compatible with such
memory devices should support a decent number of memslots (>= 509).
Message-ID: <20230926185738.277351-13-david@redhat.com>
Reviewed-by: Maciej S. Szmigiero <maciej.szmigiero@oracle.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
We want to support memory devices that have a dynamically managed memory
region container as device memory region. This device memory region maps
multiple RAM memory subregions (e.g., aliases to the same RAM memory
region), whereby these subregions can be (un)mapped on demand.
Each RAM subregion will consume a memslot in KVM and vhost, resulting in
such a new device consuming memslots dynamically, and initially usually
0. We already track the number of used vs. required memslots for all
memslots. From that, we can derive the number of reserved memslots that
must not be used otherwise.
The target use case is virtio-mem and the hyper-v balloon, which will
dynamically map aliases to RAM memory region into their device memory
region container.
Properly document what's supported and what's not and extend the vhost
memslot check accordingly.
Message-ID: <20230926185738.277351-10-david@redhat.com>
Reviewed-by: Maciej S. Szmigiero <maciej.szmigiero@oracle.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Let's track how many memslots are required by plugged memory devices and
how many are currently actually getting used by plugged memory
devices.
"required - used" is the number of reserved memslots. For now, the number
of used and required memslots is always equal, and there are no
reservations. This is a preparation for memory devices that want to
dynamically consume memslots after initially specifying how many they
require -- where we'll end up with reserved memslots.
To track the number of used memslots, create a new address space for
our device memory and register a memory listener (add/remove) for that
address space.
Message-ID: <20230926185738.277351-9-david@redhat.com>
Reviewed-by: Maciej S. Szmigiero <maciej.szmigiero@oracle.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
We want to support memory devices that have a memory region container as
device memory region that maps multiple RAM memory regions. Let's start
by supporting memory devices that statically map multiple RAM memory
regions and, thereby, consume multiple memslots.
We already have one device that uses a container as device memory region:
NVDIMMs. However, a NVDIMM always ends up consuming exactly one memslot.
Let's add support for that by asking the memory device via a new
callback how many memslots it requires.
Message-ID: <20230926185738.277351-7-david@redhat.com>
Reviewed-by: Maciej S. Szmigiero <maciej.szmigiero@oracle.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Let's return the number of free slots instead of only checking if there
is a free slot. Required to support memory devices that consume multiple
memslots.
This is a preparation for memory devices that consume multiple memslots.
Message-ID: <20230926185738.277351-6-david@redhat.com>
Reviewed-by: Maciej S. Szmigiero <maciej.szmigiero@oracle.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Let's return the number of free slots instead of only checking if there
is a free slot. While at it, check all address spaces, which will also
consider SMM under x86 correctly.
This is a preparation for memory devices that consume multiple memslots.
Message-ID: <20230926185738.277351-5-david@redhat.com>
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Reviewed-by: Maciej S. Szmigiero <maciej.szmigiero@oracle.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Checking whether the memory regions are equal is sufficient: if they are
equal, then most certainly the contained fd is equal.
The whole vhost-user memslot handling is suboptimal and overly
complicated. We shouldn't have to lookup a RAM memory regions we got
notified about in vhost_user_get_mr_data() using a host pointer. But that
requires a bigger rework -- especially an alternative vhost_set_mem_table()
backend call that simply consumes MemoryRegionSections.
For now, let's just drop vhost_backend_can_merge().
Message-ID: <20230926185738.277351-3-david@redhat.com>
Acked-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Igor Mammedov <imammedo@redhat.com>
Acked-by: Igor Mammedov <imammedo@redhat.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Having multiple vhost devices, some filtering out fd-less memslots and
some not, can mess up the "used_memslot" accounting. Consequently our
"free memslot" checks become unreliable and we might run out of free
memslots at runtime later.
An example sequence which can trigger a potential issue that involves
different vhost backends (vhost-kernel and vhost-user) and hotplugged
memory devices can be found at [1].
Let's make the filtering mechanism less generic and distinguish between
backends that support private memslots (without a fd) and ones that only
support shared memslots (with a fd). Track the used_memslots for both
cases separately and use the corresponding value when required.
Note: Most probably we should filter out MAP_PRIVATE fd-based RAM regions
(for example, via memory-backend-memfd,...,shared=off or as default with
memory-backend-file) as well. When not using MAP_SHARED, it might not work
as expected. Add a TODO for now.
[1] https://lkml.kernel.org/r/fad9136f-08d3-3fd9-71a1-502069c000cf@redhat.com
Message-ID: <20230926185738.277351-2-david@redhat.com>
Fixes: 988a27754b ("vhost: allow backends to filter memory sections")
Cc: Tiwei Bie <tiwei.bie@intel.com>
Acked-by: Igor Mammedov <imammedo@redhat.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
For migration purposes, users might want to reuse the default RAM
backend id, but specify a different memory backend.
For example, to reuse "pc.ram" on q35, one has to set
-machine q35,memory-backend=pc.ram
Only then, can a memory backend with the id "pc.ram" be created
manually.
Let's improve the error message by improving the hint. Use
error_append_hint() -- which in turn requires ERRP_GUARD().
Message-ID: <20230906120503.359863-12-david@redhat.com>
Suggested-by: ThinerLogoer <logoerthiner1@163.com>
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Tested-by: Mario Casquero <mcasquer@redhat.com>
Reviewed-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
It's easy to miss that memory-backend-file with "share=off" (default)
will always try opening the file R/W as default, and fail if we don't
have write permissions to the file.
In that case, the user has to explicit specify "readonly=on,rom=off" to
get usable RAM, for example, for VM templating.
Let's hint that '-object memory-backend-file,readonly=on,rom=off,...'
exists to consume R/O files in a private mapping to create writable RAM,
but only if we have permissions to open the file read-only.
Message-ID: <20230906120503.359863-11-david@redhat.com>
Suggested-by: ThinerLogoer <logoerthiner1@163.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Let's add some details about VM templating, focusing on the VM memory
configuration only.
There is much more to VM templating (VM state? block devices?), but I leave
that as future work.
Message-ID: <20230906120503.359863-10-david@redhat.com>
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Signed-off-by: David Hildenbrand <david@redhat.com>
"-mem-path" corresponds to "memory-backend-file,share=off" and,
therefore, creates a private COW mapping of the file. For multi-proces
QEMU, we need proper shared file-backed memory.
Let's make that clearer.
Message-ID: <20230906120503.359863-9-david@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
open() does not fail on directories when opening them readonly (O_RDONLY).
Currently, we succeed opening such directories and fail later during
mmap(), resulting in a misleading error message.
$ ./qemu-system-x86_64 \
-object memory-backend-file,id=ram0,mem-path=tmp,readonly=true,size=1g
qemu-system-x86_64: unable to map backing store for guest RAM: No such device
To identify directories and handle them accordingly in file_ram_open()
also when readonly=true was specified, detect if we just opened a directory
using fstat() instead. Then, fail file_ram_open() right away, similarly
to how we now fail if the file does not exist and we want to open the
file readonly.
With this change, we get a nicer error message:
qemu-system-x86_64: can't open backing store tmp for guest RAM: Is a directory
Note that the only memory-backend-file will end up calling
memory_region_init_ram_from_file() -> qemu_ram_alloc_from_file() ->
file_ram_open().
Message-ID: <20230906120503.359863-8-david@redhat.com>
Reported-by: Thiner Logoer <logoerthiner1@163.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Tested-by: Mario Casquero <mcasquer@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Currently, if a file does not exist yet, file_ram_open() will create new
empty file and open it writable. However, it even does that when
readonly=true was specified.
Specifying O_RDONLY instead to create a new readonly file would
theoretically work, however, ftruncate() will refuse to resize the new
empty file and we'll get a warning:
ftruncate: Invalid argument
And later eventually more problems when actually mmap'ing that file and
accessing it.
If someone intends to let QEMU open+mmap a file read-only, better
create+resize+fill that file ahead of time outside of QEMU context.
We'll now fail with:
./qemu-system-x86_64 \
-object memory-backend-file,id=ram0,mem-path=tmp,readonly=true,size=1g
qemu-system-x86_64: can't open backing store tmp for guest RAM: No such file or directory
All use cases of readonly files (R/O NVDIMMs, VM templating) work on
existing files, so silently creating new files might just hide user
errors when accidentally specifying a non-existent file.
Note that the only memory-backend-file will end up calling
memory_region_init_ram_from_file() -> qemu_ram_alloc_from_file() ->
file_ram_open().
Move error reporting to the single caller.
Message-ID: <20230906120503.359863-7-david@redhat.com>
Acked-by: Peter Xu <peterx@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
For now, "share=off,readonly=on" would always result in us opening the
file R/O and mmap'ing the opened file MAP_PRIVATE R/O -- effectively
turning it into ROM.
Especially for VM templating, "share=off" is a common use case. However,
that use case is impossible with files that lack write permissions,
because "share=off,readonly=on" will not give us writable RAM.
The sole user of ROM via memory-backend-file are R/O NVDIMMs, but as we
have users (Kata Containers) that rely on the existing behavior --
malicious VMs should not be able to consume COW memory for R/O NVDIMMs --
we cannot change the semantics of "share=off,readonly=on"
So let's add a new "rom" property with on/off/auto values. "auto" is
the default and what most people will use: for historical reasons, to not
change the old semantics, it defaults to the value of the "readonly"
property.
For VM templating, one can now use:
-object memory-backend-file,share=off,readonly=on,rom=off,...
But we'll disallow:
-object memory-backend-file,share=on,readonly=on,rom=off,...
because we would otherwise get an error when trying to mmap the R/O file
shared and writable. An explicit error message is cleaner.
We will also disallow for now:
-object memory-backend-file,share=off,readonly=off,rom=on,...
-object memory-backend-file,share=on,readonly=off,rom=on,...
It's not harmful, but also not really required for now.
Alternatives that were abandoned:
* Make "unarmed=on" for the NVDIMM set the memory region container
readonly. We would still see a change of ROM->RAM and possibly run
into memslot limits with vhost-user. Further, there might be use cases
for "unarmed=on" that should still allow writing to that memory
(temporary files, system RAM, ...).
* Add a new "readonly=on/off/auto" parameter for NVDIMMs. Similar issues
as with "unarmed=on".
* Make "readonly" consume "on/off/file" instead of being a 'bool' type.
This would slightly changes the behavior of the "readonly" parameter:
values like true/false (as accepted by a 'bool'type) would no longer be
accepted.
Message-ID: <20230906120503.359863-4-david@redhat.com>
Acked-by: Markus Armbruster <armbru@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
There is a difference between how we open a file and how we mmap it,
and we want to support writable private mappings of readonly files. Let's
define RAM_READONLY and RAM_READONLY_FD flags, to replace the single
"readonly" parameter for file-related functions.
In memory_region_init_ram_from_fd() and memory_region_init_ram_from_file(),
initialize mr->readonly based on the new RAM_READONLY flag.
While at it, add some RAM_* flags we missed to add to the list of accepted
flags in the documentation of some functions.
No change in functionality intended. We'll make use of both flags next
and start setting them independently for memory-backend-file.
Message-ID: <20230906120503.359863-3-david@redhat.com>
Acked-by: Peter Xu <peterx@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Currently, when using a true R/O NVDIMM (ROM memory backend) with a label
area, the VM can easily crash QEMU by trying to write to the label area,
because the ROM memory is mmap'ed without PROT_WRITE.
[root@vm-0 ~]# ndctl disable-region region0
disabled 1 region
[root@vm-0 ~]# ndctl zero-labels nmem0
-> QEMU segfaults
Let's remember whether we have a ROM memory backend and properly
reject the write request:
[root@vm-0 ~]# ndctl disable-region region0
disabled 1 region
[root@vm-0 ~]# ndctl zero-labels nmem0
zeroed 0 nmem
In comparison, on a system with a R/W NVDIMM:
[root@vm-0 ~]# ndctl disable-region region0
disabled 1 region
[root@vm-0 ~]# ndctl zero-labels nmem0
zeroed 1 nmem
For ACPI, just return "unsupported", like if no label exists. For spapr,
return "H_P2", similar to when no label area exists.
Could we rely on the "unarmed" property? Maybe, but it looks cleaner to
only disallow what certainly cannot work.
After all "unarmed=on" primarily means: cannot accept persistent writes. In
theory, there might be setups where devices with "unarmed=on" set could
be used to host non-persistent data (temporary files, system RAM, ...); for
example, in Linux, admins can overwrite the "readonly" setting and still
write to the device -- which will work as long as we're not using ROM.
Allowing writing label data in such configurations can make sense.
Message-ID: <20230906120503.359863-2-david@redhat.com>
Fixes: dbd730e859 ("nvdimm: check -object memory-backend-file, readonly=on option")
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Let's support device unplug by forwarding the unplug_request_check()
callback to the virtio-mem device.
Further, disallow changing the requested-size once an unplug request is
pending.
Disallowing requested-size changes handles corner cases such as
(1) pausing the VM (2) requesting device unplug and (3) adjusting the
requested size. If the VM would plug memory (due to the requested size
change) before processing the unplug request, we would be in trouble.
Message-ID: <20230711153445.514112-8-david@redhat.com>
Tested-by: Mario Casquero <mcasquer@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
In many cases, blindly unplugging a virtio-mem device is problematic. We
can only safely remove a device once:
* The guest is not expecting to be able to read unplugged memory
(unplugged-inaccessible == on)
* The virtio-mem device does not have memory plugged (size == 0)
* The virtio-mem device does not have outstanding requests to the VM to
plug memory (requested-size == 0)
So let's add a callback to the virtio-mem device class to check for that.
We'll wire-up virtio-mem-pci next.
Message-ID: <20230711153445.514112-7-david@redhat.com>
Tested-by: Mario Casquero <mcasquer@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
While we fence unplug requests from the outside, the VM can still
trigger unplug of virtio based memory devices, for example, in Linux
doing on a virtio-mem-pci device:
# echo 0 > /sys/bus/pci/slots/3/power
While doing that is not really expected to work without harming the
guest OS (e.g., removing a virtio-mem device while it still provides
memory), let's make sure that we properly handle it on the QEMU side.
We'll add support for unplugging of virtio-mem devices in some
configurations next.
Message-ID: <20230711153445.514112-5-david@redhat.com>
Tested-by: Mario Casquero <mcasquer@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Let's factor out (un)plug handling, to be reused from arm/virt code.
Provide stubs for the case that CONFIG_VIRTIO_MD is not selected because
neither virtio-mem nor virtio-pmem is enabled. While this cannot
currently happen for x86, it will be possible for arm/virt.
Message-ID: <20230711153445.514112-3-david@redhat.com>
Tested-by: Mario Casquero <mcasquer@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
To achieve desired "x-ignore-shared" functionality, we should not
discard all RAM when realizing the device and not mess with
preallocation/postcopy when loading device state. In essence, we should
not touch RAM content.
As "x-ignore-shared" gets set after realizing the device, we cannot
rely on that. Let's simply skip discarding of RAM on incoming migration.
Note that virtio_mem_post_load() will call
virtio_mem_restore_unplugged() -- unless "x-ignore-shared" is set. So
once migration finished we'll have a consistent state.
The initial system reset will also not discard any RAM, because
virtio_mem_unplug_all() will not call virtio_mem_unplug_all() when no
memory is plugged (which is the case before loading the device state).
Note that something like VM templating -- see commit b17fbbe55c
("migration: allow private destination ram with x-ignore-shared") -- is
currently incompatible with virtio-mem and ram_block_discard_range() will
warn in case a private file mapping is supplied by virtio-mem.
For VM templating with virtio-mem, it makes more sense to either
(a) Create the template without the virtio-mem device and hotplug a
virtio-mem device to the new VM instances using proper own memory
backend.
(b) Use a virtio-mem device that doesn't provide any memory in the
template (requested-size=0) and use private anonymous memory.
Message-ID: <20230706075612.67404-5-david@redhat.com>
Tested-by: Mario Casquero <mcasquer@redhat.com>
Reviewed-by: Juan Quintela <quintela@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
virtio-mem wants to know whether it should not mess with the RAMBlock
content (e.g., discard RAM, preallocate memory) on incoming migration.
So let's expose that function as migrate_ram_is_ignored() in
migration/misc.h
Message-ID: <20230706075612.67404-4-david@redhat.com>
Acked-by: Peter Xu <peterx@redhat.com>
Tested-by: Mario Casquero <mcasquer@redhat.com>
Reviewed-by: Juan Quintela <quintela@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Already when starting QEMU we perform one system reset that ends up
triggering virtio_mem_unplug_all() with no actual memory plugged yet.
That, in turn will trigger ram_block_discard_range() and perform some
other actions that are not required in that case.
Let's optimize virtio_mem_unplug_all() for the case that no memory is
plugged. This will be beneficial for x-ignore-shared support as well.
Message-ID: <20230706075612.67404-3-david@redhat.com>
Tested-by: Mario Casquero <mcasquer@redhat.com>
Reviewed-by: Juan Quintela <quintela@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
ram_block_discard_range() cannot possibly do the right thing in
MAP_PRIVATE file mappings in the general case.
To achieve the documented semantics, we also have to punch a hole into
the file, possibly messing with other MAP_PRIVATE/MAP_SHARED mappings
of such a file.
For example, using VM templating -- see commit b17fbbe55c ("migration:
allow private destination ram with x-ignore-shared") -- in combination with
any mechanism that relies on discarding of RAM is problematic. This
includes:
* Postcopy live migration
* virtio-balloon inflation/deflation or free-page-reporting
* virtio-mem
So at least warn that there is something possibly dangerous is going on
when using ram_block_discard_range() in these cases.
Message-ID: <20230706075612.67404-2-david@redhat.com>
Acked-by: Peter Xu <peterx@redhat.com>
Tested-by: Mario Casquero <mcasquer@redhat.com>
Reviewed-by: Juan Quintela <quintela@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Let's move memory_device_check_addable() and basic checks out of
memory_device_get_free_addr() directly into memory_device_pre_plug().
Separating basic checks from address assignment is cleaner and
prepares for further changes.
As all memory device users now use memory_devices_init(), and that
function enforces that the size is 0, we can drop the check for an empty
region.
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Signed-off-by: David Hildenbrand <david@redhat.com>
Message-Id: <20230623124553.400585-10-david@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Let's use our new helper and stop always allocating ms->device_memory.
There is no difference in common memory-device code anymore between
ms->device_memory being NULL or the size being 0. So we only have to
teach spapr code that ms->device_memory isn't always around.
We can now modify two maxram_size checks to rely on ms->device_memory
for detecting whether we have memory devices.
Cc: Daniel Henrique Barboza <danielhb413@gmail.com>
Cc: "Cédric Le Goater" <clg@kaod.org>
Cc: David Gibson <david@gibson.dropbear.id.au>
Cc: Greg Kurz <groug@kaod.org>
Cc: Harsh Prateek Bora <harshpb@linux.ibm.com>
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Reviewed-by: Cédric Le Goater <clg@kaod.org>
Signed-off-by: David Hildenbrand <david@redhat.com>
Message-Id: <20230623124553.400585-5-david@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Let's intrduce a new helper that we will use to replace existing memory
device setup code during machine initialization. We'll enforce that the
size has to be > 0.
Once all machines were converted, we'll only allocate ms->device_memory
if the size > 0.
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Signed-off-by: David Hildenbrand <david@redhat.com>
Message-Id: <20230623124553.400585-3-david@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Let's unify the error messages, such that we can simply stop allocating
ms->device_memory if the size would be 0 (and there are no memory
devices ever).
The case of "not supported by the machine" should barely pop up either
way: if the machine doesn't support memory devices, it usually doesn't
call the pre_plug handler ...
Reviewed-by: Philippe Mathieu-Daudé <philmd@linaro.org>
Signed-off-by: David Hildenbrand <david@redhat.com>
Message-Id: <20230623124553.400585-2-david@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Let's separate plug and unplug handling to prepare for future changes
and make the code a bit easier to read -- working on block states
(plugged/unplugged) instead of on a bitmap.
Cc: "Michael S. Tsirkin" <mst@redhat.com>
Cc: Gavin Shan <gshan@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Message-Id: <20230523183036.517957-1-david@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Let's just support 512 memslots on x86-64 and aarch64 as well. The maximum
number of ACPI slots (256) is no longer completely expressive ever since
we supported virtio-based memory devices. Further, we're completely
ignoring other memslots used outside of memory device context, such as
memslots used for boot memory.
Note that the vhost memslot limit in the kernel is usually configured to
be 509. With this change, we prepare vhost-user on the QEMU side to be
closer to that limit, to eventually support ~512 memslots in most vhost
implementations and have less "surprises" when cold/hotplugging vhost
devices while also consuming more memslots than we're currently used to
by memory devices (e.g., once virtio-mem starts using multiple memslots).
Note that most vhost-user implementations only support a small number of
memslots so far, which we can hopefully improve in the near future.
We'll leave the PPC special-case as is for now.
Cc: "Michael S. Tsirkin" <mst@redhat.com>
Cc: Igor Mammedov <imammedo@redhat.com>
Cc: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Message-Id: <20230503184144.808478-1-david@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Allowing guests to read unplugged memory simplified the bring-up of
virtio-mem in Linux guests -- which was limited to x86-64 only. On arm64
(which was added later), we never had legacy guests and don't even allow
to configure it, essentially always having "unplugged-inaccessible=on".
At this point, all guests we care about
should be supporting VIRTIO_MEM_F_UNPLUGGED_INACCESSIBLE, so let's
change the default for the 8.1 machine.
This change implies that also memory that supports the shared zeropage
(private anonymous memory) will now require
VIRTIO_MEM_F_UNPLUGGED_INACCESSIBLE in the driver in order to be usable by
the guest -- as default, one can still manually set the
unplugged-inaccessible property.
Disallowing the guest to read unplugged memory will be important for
some future features, such as memslot optimizations or protection of
unplugged memory, whereby we'll actually no longer allow the guest to
even read from unplugged memory.
At some point, we might want to deprecate and remove that property.
Cc: "Michael S. Tsirkin" <mst@redhat.com>
Cc: Marcel Apfelbaum <marcel.apfelbaum@gmail.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Richard Henderson <richard.henderson@linaro.org>
Cc: Eduardo Habkost <eduardo@habkost.net>
Signed-off-by: David Hildenbrand <david@redhat.com>
Message-Id: <20230503182352.792458-1-david@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Ordinary memory preallocation runs when QEMU starts up and creates the
memory backends, before processing the incoming migration stream. With
virtio-mem, we don't know which memory blocks to preallocate before
migration started. Now that we migrate the virtio-mem bitmap early, before
migrating any RAM content, we can safely preallocate memory for all plugged
memory blocks before migrating any RAM content.
This is especially relevant for the following cases:
(1) User errors
With hugetlb/files, if we don't have sufficient backend memory available on
the migration destination, we'll crash QEMU (SIGBUS) during RAM migration
when running out of backend memory. Preallocating memory before actual
RAM migration allows for failing gracefully and informing the user about
the setup problem.
(2) Excluded memory ranges during migration
For example, virtio-balloon free page hinting will exclude some pages
from getting migrated. In that case, we won't crash during RAM
migration, but later, when running the VM on the destination, which is
bad.
To fix this for new QEMU machines that migrate the bitmap early,
preallocate the memory early, before any RAM migration. Warn with old
QEMU machines.
Getting postcopy right is a bit tricky, but we essentially now implement
the same (problematic) preallocation logic as ordinary preallocation:
preallocate memory early and discard it again before precopy starts. During
ordinary preallocation, discarding of RAM happens when postcopy is advised.
As the state (bitmap) is loaded after postcopy was advised but before
postcopy starts listening, we have to discard memory we preallocated
immediately again ourselves.
Note that nothing (not even hugetlb reservations) guarantees for postcopy
that backend memory (especially, hugetlb pages) are still free after they
were freed ones while discarding RAM. Still, allocating that memory at
least once helps catching some basic setup problems.
Before this change, trying to restore a VM when insufficient hugetlb
pages are around results in the process crashing to to a "Bus error"
(SIGBUS). With this change, QEMU fails gracefully:
qemu-system-x86_64: qemu_prealloc_mem: preallocating memory failed: Bad address
qemu-system-x86_64: error while loading state for instance 0x0 of device '0000:00:03.0/virtio-mem-device-early'
qemu-system-x86_64: load of migration failed: Cannot allocate memory
And we can even introspect the early migration data, including the
bitmap:
$ ./scripts/analyze-migration.py -f STATEFILE
{
"ram (2)": {
"section sizes": {
"0000:00:03.0/mem0": "0x0000000780000000",
"0000:00:04.0/mem1": "0x0000000780000000",
"pc.ram": "0x0000000100000000",
"/rom@etc/acpi/tables": "0x0000000000020000",
"pc.bios": "0x0000000000040000",
"0000:00:02.0/e1000.rom": "0x0000000000040000",
"pc.rom": "0x0000000000020000",
"/rom@etc/table-loader": "0x0000000000001000",
"/rom@etc/acpi/rsdp": "0x0000000000001000"
}
},
"0000:00:03.0/virtio-mem-device-early (51)": {
"tmp": "00 00 00 01 40 00 00 00 00 00 00 07 80 00 00 00 00 00 00 00 00 20 00 00 00 00 00 00",
"size": "0x0000000040000000",
"bitmap": "ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff [...]
},
"0000:00:04.0/virtio-mem-device-early (53)": {
"tmp": "00 00 00 08 c0 00 00 00 00 00 00 07 80 00 00 00 00 00 00 00 00 20 00 00 00 00 00 00",
"size": "0x00000001fa400000",
"bitmap": "ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff ff [...]
},
[...]
Reported-by: Jing Qi <jinqi@redhat.com>
Reviewed-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Reviewed-by: Juan Quintela <quintela@redhat.com>S
Signed-off-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Juan Quintela <quintela@redhat.com>
The bitmap and the size are immutable while migration is active: see
virtio_mem_is_busy(). We can migrate this information early, before
migrating any actual RAM content. Further, all information we need for
sanity checks is immutable as well.
Having this information in place early will, for example, allow for
properly preallocating memory before touching these memory locations
during RAM migration: this way, we can make sure that all memory was
actually preallocated and that any user errors (e.g., insufficient
hugetlb pages) can be handled gracefully.
In contrast, usable_region_size and requested_size can theoretically
still be modified on the source while the VM is running. Keep migrating
these properties the usual, late, way.
Use a new device property to keep behavior of compat machines
unmodified.
Reviewed-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Reviewed-by: Juan Quintela <quintela@redhat.com>S
Signed-off-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Juan Quintela <quintela@redhat.com>
"prealloc=on" for the memory backend does not work as expected, as
virtio-mem will simply discard all preallocated memory immediately again.
In the best case, it's an expensive NOP. In the worst case, it's an
unexpected allocation error.
Instead, "prealloc=on" should be specified for the virtio-mem device only,
such that virtio-mem will try preallocating memory before plugging
memory dynamically to the guest. Fail if such a memory backend is
provided.
Tested-by: Michal Privoznik <mprivozn@redhat.com>
Reviewed-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Reviewed-by: Juan Quintela <quintela@redhat.com>S
Signed-off-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Juan Quintela <quintela@redhat.com>
Let's factor out this check, to be used in virtio-mem context next.
While at it, fix a spelling error in a related comment.
Reviewed-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Reviewed-by: Juan Quintela <quintela@redhat.com>S
Signed-off-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Juan Quintela <quintela@redhat.com>
For virtio-mem, we want to have the plugged/unplugged state of memory
blocks available before migrating any actual RAM content, and perform
sanity checks before touching anything on the destination. This
information is immutable on the migration source while migration is active,
We want to use this information for proper preallocation support with
migration: currently, we don't preallocate memory on the migration target,
and especially with hugetlb, we can easily run out of hugetlb pages during
RAM migration and will crash (SIGBUS) instead of catching this gracefully
via preallocation.
Migrating device state via a VMSD before we start iterating is currently
impossible: the only approach that would be possible is avoiding a VMSD
and migrating state manually during save_setup(), to be restored during
load_state().
Let's allow for migrating device state via a VMSD early, during the
setup phase in qemu_savevm_state_setup(). To keep it simple, we
indicate applicable VMSD's using an "early_setup" flag.
Note that only very selected devices (i.e., ones seriously messing with
RAM setup) are supposed to make use of such early state migration.
While at it, also use a bool for the "unmigratable" member.
Reviewed-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Reviewed-by: Juan Quintela <quintela@redhat.com>S
Signed-off-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Juan Quintela <quintela@redhat.com>
... and store it in the migration state. This is a preparation for
storing selected vmds's already in qemu_savevm_state_setup().
Reviewed-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Reviewed-by: Juan Quintela <quintela@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Juan Quintela <quintela@redhat.com>
Let's move more code into vmstate_save(), reducing code duplication and
preparing for reuse of vmstate_save() in qemu_savevm_state_setup(). We
have to move vmstate_save() to make the compiler happy.
We'll now also trace from qemu_save_device_state(), triggering the same
tracepoints as previously called from
qemu_savevm_state_complete_precopy_non_iterable() only. Note that
qemu_save_device_state() ignores iterable device state, such as RAM,
and consequently doesn't trigger some other trace points (e.g.,
trace_savevm_state_setup()).
Reviewed-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Reviewed-by: Juan Quintela <quintela@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Juan Quintela <quintela@redhat.com>
ram_block_populate_read() already optimizes for RamDiscardManager.
However, ram_write_tracking_start() will still try protecting discarded
memory ranges.
Let's optimize, because discarded ranges don't map any pages and
(1) For anonymous memory, trying to protect using uffd-wp without a mapped
page is ignored by the kernel and consequently a NOP.
(2) For shared/file-backed memory, we will fill present page tables in the
range with PTE markers. However, we will even allocate page tables
just to fill them with unnecessary PTE markers and effectively
waste memory.
So let's exclude these ranges, just like ram_block_populate_read()
already does.
Reviewed-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Juan Quintela <quintela@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Juan Quintela <quintela@redhat.com>
ram_mig_ram_block_resized() will abort migration (including background
snapshots) when resizing a RAMBlock. ram_block_populate_read() will only
populate RAM up to used_length, so at least for anonymous memory
protecting everything between used_length and max_length won't
actually be protected and is just a NOP.
So let's only protect everything up to used_length.
Note: it still makes sense to register uffd-wp for max_length, such
that RAM_UF_WRITEPROTECT is independent of a changing used_length.
Reviewed-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Juan Quintela <quintela@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Juan Quintela <quintela@redhat.com>
When unregistering uffd-wp, older kernels before commit f369b07c86143
("mm/uffd:reset write protection when unregister with wp-mode") won't
clear the uffd-wp PTE bit. When re-registering uffd-wp, the previous
uffd-wp PTE bits would trigger again. With above commit, the kernel will
clear the uffd-wp PTE bits when unregistering itself.
Consequently, we'll clear the uffd-wp PTE bits now twice -- whereby we
don't care about clearing them at all: a new background snapshot will
re-register uffd-wp and re-protect all memory either way.
So let's skip the manual clearing of uffd-wp. If ever relevant, we
could clear conditionally in uffd_unregister_memory() -- we just need a
way to figure out more recent kernels.
Reviewed-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Juan Quintela <quintela@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Juan Quintela <quintela@redhat.com>
If something goes wrong during uffd_change_protection(), we would miss
to unregister uffd-wp and not release our reference. Fix it by
performing the uffd_change_protection(true) last.
Note that a uffd_change_protection(false) on the recovery path without a
prior uffd_change_protection(false) is fine.
Fixes: 278e2f551a ("migration: support UFFD write fault processing in ram_save_iterate()")
Cc: qemu-stable@nongnu.org
Reviewed-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Juan Quintela <quintela@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Juan Quintela <quintela@redhat.com>
Unfortunately, commit f7b9dcfbcf broke populate_read_range(): the loop
end condition is very wrong, resulting in that function not populating the
full range. Lets' fix that.
Fixes: f7b9dcfbcf ("migration/ram: Factor out populating pages readable in ram_block_populate_pages()")
Cc: qemu-stable@nongnu.org
Reviewed-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Juan Quintela <quintela@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Juan Quintela <quintela@redhat.com>
If we update an existing memslot (e.g., resize, split), we temporarily
remove the memslot to re-add it immediately afterwards. These updates
are not atomic, especially not for KVM VCPU threads, such that we can
get spurious faults.
Let's inhibit most KVM ioctls while performing relevant updates, such
that we can perform the update just as if it would happen atomically
without additional kernel support.
We capture the add/del changes and apply them in the notifier commit
stage instead. There, we can check for overlaps and perform the ioctl
inhibiting only if really required (-> overlap).
To keep things simple we don't perform additional checks that wouldn't
actually result in an overlap -- such as !RAM memory regions in some
cases (see kvm_set_phys_mem()).
To minimize cache-line bouncing, use a separate indicator
(in_ioctl_lock) per CPU. Also, make sure to hold the kvm_slots_lock
while performing both actions (removing+re-adding).
We have to wait until all IOCTLs were exited and block new ones from
getting executed.
This approach cannot result in a deadlock as long as the inhibitor does
not hold any locks that might hinder an IOCTL from getting finished and
exited - something fairly unusual. The inhibitor will always hold the BQL.
AFAIKs, one possible candidate would be userfaultfd. If a page cannot be
placed (e.g., during postcopy), because we're waiting for a lock, or if the
userfaultfd thread cannot process a fault, because it is waiting for a
lock, there could be a deadlock. However, the BQL is not applicable here,
because any other guest memory access while holding the BQL would already
result in a deadlock.
Nothing else in the kernel should block forever and wait for userspace
intervention.
Note: pause_all_vcpus()/resume_all_vcpus() or
start_exclusive()/end_exclusive() cannot be used, as they either drop
the BQL or require to be called without the BQL - something inhibitors
cannot handle. We need a low-level locking mechanism that is
deadlock-free even when not releasing the BQL.
Signed-off-by: David Hildenbrand <david@redhat.com>
Signed-off-by: Emanuele Giuseppe Esposito <eesposit@redhat.com>
Tested-by: Emanuele Giuseppe Esposito <eesposit@redhat.com>
Message-Id: <20221111154758.1372674-4-eesposit@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Currently, there is no way to configure a CPU affinity inside QEMU when
the sandbox option disables it for QEMU as a whole, for example, via:
-sandbox enable=on,resourcecontrol=deny
While ThreadContext objects can be created on the QEMU commandline and
the CPU affinity can be configured externally via the thread-id, this is
insufficient if a ThreadContext with a certain CPU affinity is already
required during QEMU startup, before we can intercept QEMU and
configure the CPU affinity.
Blocking sched_setaffinity() was introduced in 24f8cdc572 ("seccomp:
add resourcecontrol argument to command line"), "to avoid any bigger of the
process". However, we only care about once QEMU is running, not when
the instance starting QEMU explicitly requests a certain CPU affinity
on the QEMU comandline.
Right now, for NUMA-aware preallocation of memory backends used for initial
machine RAM, one has to:
1) Start QEMU with the memory-backend with "prealloc=off"
2) Pause QEMU before it starts the guest (-S)
3) Create ThreadContext, configure the CPU affinity using the thread-id
4) Configure the ThreadContext as "prealloc-context" of the memory
backend
5) Trigger preallocation by setting "prealloc=on"
To simplify this handling especially for initial machine RAM,
allow creation of ThreadContext objects before parsing sandbox options,
such that the CPU affinity requested on the QEMU commandline alongside the
sandbox option can be set. As ThreadContext objects essentially only create
a persistent context thread and set the CPU affinity, this is easily
possible.
With this change, we can create a ThreadContext with a CPU affinity on
the QEMU commandline and use it for preallocation of memory backends
glued to the machine (simplified example):
To make "-name debug-threads=on" keep working as expected for the
context threads, perform earlier parsing of "-name".
qemu-system-x86_64 -m 1G \
-object thread-context,id=tc1,cpu-affinity=3-4 \
-object memory-backend-ram,id=pc.ram,size=1G,prealloc=on,prealloc-threads=2,prealloc-context=tc1 \
-machine memory-backend=pc.ram \
-S -monitor stdio -sandbox enable=on,resourcecontrol=deny
And while we can query the current CPU affinity:
(qemu) qom-get tc1 cpu-affinity
[
3,
4
]
We can no longer change it from QEMU directly:
(qemu) qom-set tc1 cpu-affinity 1-2
Error: Setting CPU affinity failed: Operation not permitted
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
Message-Id: <20221014134720.168738-8-david@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Let's allow for specifying a thread context via the "prealloc-context"
property. When set, preallcoation threads will be crated via the
thread context -- inheriting the same CPU affinity as the thread
context.
Pinning preallcoation threads to CPUs can heavily increase performance
in NUMA setups, because, preallocation from a CPU close to the target
NUMA node(s) is faster then preallocation from a CPU further remote,
simply because of memory bandwidth for initializing memory with zeroes.
This is especially relevant for very large VMs backed by huge/gigantic
pages, whereby preallocation is mandatory.
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
Message-Id: <20221014134720.168738-7-david@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
... and implement it under POSIX. When a ThreadContext is provided,
create new threads via the context such that these new threads obtain a
properly configured CPU affinity.
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
Message-Id: <20221014134720.168738-6-david@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Let's make it easier to pin threads created via a ThreadContext to
all host CPUs currently belonging to a given set of host NUMA nodes --
which is the common case.
"node-affinity" is simply a shortcut for setting "cpu-affinity" manually
to the list of host CPUs belonging to the set of host nodes. This property
can only be written.
A simple QEMU example to set the CPU affinity to host node 1 on a system
with two nodes, 24 CPUs each, whereby odd-numbered host CPUs belong to
host node 1:
qemu-system-x86_64 -S \
-object thread-context,id=tc1,node-affinity=1
And we can query the cpu-affinity via HMP/QMP:
(qemu) qom-get tc1 cpu-affinity
[
1,
3,
5,
7,
9,
11,
13,
15,
17,
19,
21,
23,
25,
27,
29,
31,
33,
35,
37,
39,
41,
43,
45,
47
]
We cannot query the node-affinity:
(qemu) qom-get tc1 node-affinity
Error: Insufficient permission to perform this operation
But note that due to dynamic library loading this example will not work
before we actually make use of thread_context_create_thread() in QEMU
code, because the type will otherwise not get registered. We'll wire
this up next to make it work.
Note that if the host CPUs for a host node change due do CPU hot(un)plug
CPU onlining/offlining (i.e., lscpu output changes) after the ThreadContext
was started, the CPU affinity will not get updated.
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
Acked-by: Markus Armbruster <armbru@redhat.com>
Message-Id: <20221014134720.168738-5-david@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Setting the CPU affinity of QEMU threads is a bit problematic, because
QEMU doesn't always have permissions to set the CPU affinity itself,
for example, with seccomp after initialized by QEMU:
-sandbox enable=on,resourcecontrol=deny
General information about CPU affinities can be found in the man page of
taskset:
CPU affinity is a scheduler property that "bonds" a process to a given
set of CPUs on the system. The Linux scheduler will honor the given CPU
affinity and the process will not run on any other CPUs.
While upper layers are already aware of how to handle CPU affinities for
long-lived threads like iothreads or vcpu threads, especially short-lived
threads, as used for memory-backend preallocation, are more involved to
handle. These threads are created on demand and upper layers are not even
able to identify and configure them.
Introduce the concept of a ThreadContext, that is essentially a thread
used for creating new threads. All threads created via that context
thread inherit the configured CPU affinity. Consequently, it's
sufficient to create a ThreadContext and configure it once, and have all
threads created via that ThreadContext inherit the same CPU affinity.
The CPU affinity of a ThreadContext can be configured two ways:
(1) Obtaining the thread id via the "thread-id" property and setting the
CPU affinity manually (e.g., via taskset).
(2) Setting the "cpu-affinity" property and letting QEMU try set the
CPU affinity itself. This will fail if QEMU doesn't have permissions
to do so anymore after seccomp was initialized.
A simple QEMU example to set the CPU affinity to host CPU 0,1,6,7 would be:
qemu-system-x86_64 -S \
-object thread-context,id=tc1,cpu-affinity=0-1,cpu-affinity=6-7
And we can query it via HMP/QMP:
(qemu) qom-get tc1 cpu-affinity
[
0,
1,
6,
7
]
But note that due to dynamic library loading this example will not work
before we actually make use of thread_context_create_thread() in QEMU
code, because the type will otherwise not get registered. We'll wire
this up next to make it work.
In general, the interface behaves like pthread_setaffinity_np(): host
CPU numbers that are currently not available are ignored; only host CPU
numbers that are impossible with the current kernel will fail. If the
list of host CPU numbers does not include a single CPU that is
available, setting the CPU affinity will fail.
A ThreadContext can be reused, simply by reconfiguring the CPU affinity.
Note that the CPU affinity of previously created threads will not get
adjusted.
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
Acked-by: Markus Armbruster <armbru@redhat.com>
Message-Id: <20221014134720.168738-4-david@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Usually, we let upper layers handle CPU pinning, because
pthread_setaffinity_np() (-> sched_setaffinity()) is blocked via
seccomp when starting QEMU with
-sandbox enable=on,resourcecontrol=deny
However, we want to configure and observe the CPU affinity of threads
from QEMU directly in some cases when the sandbox option is either not
enabled or not active yet.
So let's add a way to configure CPU pinning via
qemu_thread_set_affinity() and obtain CPU affinity via
qemu_thread_get_affinity() and implement them under POSIX using
pthread_setaffinity_np() + pthread_getaffinity_np().
Implementation under Windows is possible using SetProcessAffinityMask()
+ GetProcessAffinityMask(), however, that is left as future work.
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
Message-Id: <20221014134720.168738-3-david@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Let's
* give the function a "qemu_*" style name
* make sure the parameters in the implementation match the prototype
* rename smp_cpus to max_threads, which makes the semantics of that
parameter clearer
... and add a function documentation.
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
Message-Id: <20221014134720.168738-2-david@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Xiao Guangrong doesn't have enough time to actively review or contribute
to our NVDIMM implementation. Let's dissolve the "NVDIMM" section, moving
relevant ACPI parts to "ACPI/SMBIOS" and moving memory device stuff into a
new "Memory devices" section. Make that new section cover other memory
device stuff as well.
We can now drop the "hw/mem/*" rule from "ACPI/SMBIOS". Note that
hw/acpi/nvdimm.c is already covered by "ACPI/SMBIOS".
The following files in hw/mem don't fall into the TYPE_MEMPORY_DEVICE
category:
* hw/mem/cxl_type3.c is CXL specific and belongs to "Compute Express Link"
* hw/mem/sparse-mem.c is already covered by "Device Fuzzing"
* hw/mem/npcm7xx_mc.c is already covered by "Nuvoton NPCM7xx"
Thanks Xiao for your work on NVDIMM!
Cc: Ben Widawsky <ben.widawsky@intel.com>
Cc: Jonathan Cameron <jonathan.cameron@huawei.com>
Cc: Michael S. Tsirkin <mst@redhat.com>
Cc: Igor Mammedov <imammedo@redhat.com>
Cc: Ani Sinha <ani@anisinha.ca>
Cc: Xiao Guangrong <xiaoguangrong.eric@gmail.com>
Cc: "Philippe Mathieu-Daudé" <f4bug@amsat.org>
Cc: Richard Henderson <richard.henderson@linaro.org>
Cc: Peter Maydell <peter.maydell@linaro.org>
Cc: Julia Suvorova <jusual@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Message-Id: <20220617123151.103033-1-david@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Before we were able to bump up the qemu CPU model to a z13, we included
some experimental features during development in the "max" model only.
Nowadays, the "max" model corresponds exactly to the "qemu" CPU model
of the latest QEMU machine under TCG.
Let's remove all the special casing, effectively making both models
match completely from now on, and clean up.
Signed-off-by: David Hildenbrand <david@redhat.com>
Reviewed-by: Cornelia Huck <cohuck@redhat.com>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Message-Id: <20220428094708.84835-4-david@redhat.com>
Signed-off-by: Thomas Huth <thuth@redhat.com>
We don't include the "msa5" feature in the "qemu" model because it
generates a warning. The PoP states:
"The message-security-assist extension 5 requires
the secure-hash-algorithm (SHA-512) capabilities of
the message-security-assist extension 2 as a prereq-
uisite. (March, 2015)"
As SHA-512 won't be supported in the near future, let's just drop the
feature from the "max" model. This avoids the warning and allows us for
making the "max" model match the "qemu" model (except for compat
machines). We don't lose much, as we only implement the function stubs
for MSA, excluding any real subfunctions.
Signed-off-by: David Hildenbrand <david@redhat.com>
Reviewed-by: Thomas Huth <thuth@redhat.com>
Reviewed-by: Cornelia Huck <cohuck@redhat.com>
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Resolves: https://gitlab.com/qemu-project/qemu/-/issues/897
Message-Id: <20220428094708.84835-3-david@redhat.com>
Signed-off-by: Thomas Huth <thuth@redhat.com>
memory_region_readd_subregion() wants to readd a region by first
removing it and then readding it. For readding, it doesn't use one of
the memory_region_add_*() variants, which is why fail to re-increment the
mr->mapped_via_alias counters, resulting in the
assert(alias->mapped_via_alias >= 0) in memory_region_del_subregion()
triggering the next time we call memory_region_readd_subregion().
Fix it by using memory_region_add_subregion_common() for readding the
region.
Reported-by: Niek Linnenbank <nieklinnenbank@gmail.com>
Fixes: 5ead62185d ("memory: Make memory_region_is_mapped() succeed when mapped via an alias")
Tested-by: Niek Linnenbank <nieklinnenbank@gmail.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: "Philippe Mathieu-Daudé" <f4bug@amsat.org>
Signed-off-by: David Hildenbrand <david@redhat.com>
Message-Id: <20220201100940.47788-1-david@redhat.com>
Signed-off-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
For fd-based shared memory, MAP_NORESERVE is only effective for hugetlb,
otherwise it's ignored. Older Linux versions that didn't support
reservation of huge pages ignored MAP_NORESERVE completely.
The first client to mmap a hugetlb fd without MAP_NORESERVE will
trigger reservation of huge pages for the whole mmapped range. There are
two cases to consider:
1) QEMU mapped RAM without MAP_NORESERVE
We're not dealing with a sparse mapping, huge pages for the whole range
have already been reserved by QEMU. An additional mmap() without
MAP_NORESERVE won't have any effect on the reservation.
2) QEMU mapped RAM with MAP_NORESERVE
We're delaing with a sparse mapping, no huge pages should be reserved.
Further mappings without MAP_NORESERVE should be avoided.
For 1), it doesn't matter if we set MAP_NORESERVE or not, so we can
simply set it. For 2), we'd be overriding QEMUs decision and trigger
reservation of huge pages, which might just fail if there are not
sufficient huge pages around. We must map with MAP_NORESERVE.
This change is required to support virtio-mem with hugetlb: a
virtio-mem device mapped into the guest physical memory corresponds to
a sparse memory mapping and QEMU maps this memory with MAP_NORESERVE.
Whenever memory in that sparse region will be accessed by the VM, QEMU
populates huge pages for the affected range by preallocating memory
and handling any preallocation errors gracefully.
So let's map shared RAM with MAP_NORESERVE. As libvhost-user only
supports Linux, there shouldn't be anything to take care of in regard of
other OS support.
Without this change, libvhost-user will fail mapping the region if there
are currently not enough huge pages to perform the reservation:
fv_panic: libvhost-user: region mmap error: Cannot allocate memory
Cc: "Marc-André Lureau" <marcandre.lureau@redhat.com>
Cc: "Michael S. Tsirkin" <mst@redhat.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Raphael Norwitz <raphael.norwitz@nutanix.com>
Cc: Stefan Hajnoczi <stefanha@redhat.com>
Cc: Dr. David Alan Gilbert <dgilbert@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Message-Id: <20220111123939.132659-1-david@redhat.com>
Acked-by: Raphael Norwitz <raphael.norwitz@nutanix.com>
memory_region_is_mapped() currently does not return "true" when a memory
region is mapped via an alias.
Assuming we have:
alias (A0) -> alias (A1) -> region (R0)
Mapping A0 would currently only make memory_region_is_mapped() succeed
on A0, but not on A1 and R0.
Let's fix that by adding a "mapped_via_alias" counter to memory regions and
updating it accordingly when an alias gets (un)mapped.
I am not aware of actual issues, this is rather a cleanup to make it
consistent.
Reviewed-by: Philippe Mathieu-Daudé <philmd@redhat.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Message-Id: <20211102164317.45658-3-david@redhat.com>
Signed-off-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
memory_region_is_mapped() is the wrong check, we actually want to check
whether the backend is already marked mapped.
For example, memory regions mapped via an alias, such as NVDIMMs,
currently don't make memory_region_is_mapped() return "true". As the
machine is initialized before any memory devices (and thereby before
NVDIMMs are initialized), this isn't a fix but merely a cleanup.
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Igor Mammedov <imammedo@redhat.com>
Reviewed-by: Philippe Mathieu-Daudé <philmd@redhat.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Message-Id: <20211102164317.45658-2-david@redhat.com>
Signed-off-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
With VIRTIO_MEM_F_UNPLUGGED_INACCESSIBLE, we signal the VM that reading
unplugged memory is not supported. We have to fail feature negotiation
in case the guest does not support VIRTIO_MEM_F_UNPLUGGED_INACCESSIBLE.
First, VIRTIO_MEM_F_UNPLUGGED_INACCESSIBLE is required to properly handle
memory backends (or architectures) without support for the shared zeropage
in the hypervisor cleanly. Without the shared zeropage, even reading an
unpopulated virtual memory location can populate real memory and
consequently consume memory in the hypervisor. We have a guaranteed shared
zeropage only on MAP_PRIVATE anonymous memory.
Second, we want VIRTIO_MEM_F_UNPLUGGED_INACCESSIBLE to be the default
long-term as even populating the shared zeropage can be problematic: for
example, without THP support (possible) or without support for the shared
huge zeropage with THP (unlikely), the PTE page tables to hold the shared
zeropage entries can consume quite some memory that cannot be reclaimed
easily.
Third, there are other optimizations+features (e.g., protection of
unplugged memory, reducing the total memory slot size and bitmap sizes)
that will require VIRTIO_MEM_F_UNPLUGGED_INACCESSIBLE.
We really only support x86 targets with virtio-mem for now (and
Linux similarly only support x86), but that might change soon, so prepare
for different targets already.
Add a new "unplugged-inaccessible" tristate property for x86 targets:
- "off" will keep VIRTIO_MEM_F_UNPLUGGED_INACCESSIBLE unset and legacy
guests working.
- "on" will set VIRTIO_MEM_F_UNPLUGGED_INACCESSIBLE and stop legacy guests
from using the device.
- "auto" selects the default based on support for the shared zeropage.
Warn in case the property is set to "off" and we don't have support for the
shared zeropage.
For existing compat machines, the property will default to "off", to
not change the behavior but eventually warn about a problematic setup.
Short-term, we'll set the property default to "auto" for new QEMU machines.
Mid-term, we'll set the property default to "on" for new QEMU machines.
Long-term, we'll deprecate the parameter and disallow legacy
guests completely.
The property has to match on the migration source and destination. "auto"
will result in the same VIRTIO_MEM_F_UNPLUGGED_INACCESSIBLE setting as long
as the qemu command line (esp. memdev) match -- so "auto" is good enough
for migration purposes and the parameter doesn't have to be migrated
explicitly.
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Message-Id: <20211217134039.29670-3-david@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
For scarce memory resources, such as hugetlb, we want to be able to
prealloc such memory resources in order to not crash later on access. On
simple user errors we could otherwise easily run out of memory resources
an crash the VM -- pretty much undesired.
For ordinary memory devices, such as DIMMs, we preallocate memory via the
memory backend for such use cases; however, with virtio-mem we're dealing
with sparse memory backends; preallocating the whole memory backend
destroys the whole purpose of virtio-mem.
Instead, we want to preallocate memory when actually exposing memory to the
VM dynamically, and fail plugging memory gracefully + warn the user in case
preallocation fails.
A common use case for hugetlb will be using "reserve=off,prealloc=off" for
the memory backend and "prealloc=on" for the virtio-mem device. This
way, no huge pages will be reserved for the process, but we can recover
if there are no actual huge pages when plugging memory. Libvirt is
already prepared for this.
Note that preallocation cannot protect from the OOM killer -- which
holds true for any kind of preallocation in QEMU. It's primarily useful
only for scarce memory resources such as hugetlb, or shared file-backed
memory. It's of little use for ordinary anonymous memory that can be
swapped, KSM merged, ... but we won't forbid it.
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Message-Id: <20211217134611.31172-9-david@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Temporarily modifying the SIGBUS handler is really nasty, as we might be
unlucky and receive an MCE SIGBUS while having our handler registered.
Unfortunately, there is no way around messing with SIGBUS when
MADV_POPULATE_WRITE is not applicable or not around.
Let's forward SIGBUS that don't belong to us to the already registered
handler and document the situation.
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Message-Id: <20211217134611.31172-8-david@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Add a mutex to protect the SIGBUS case, as we cannot mess concurrently
with the sigbus handler and we have to manage the global variable
sigbus_memset_context. The MADV_POPULATE_WRITE path can run
concurrently.
Note that page_mutex and page_cond are shared between concurrent
invocations, which shouldn't be a problem.
This is a preparation for future virtio-mem prealloc code, which will call
os_mem_prealloc() asynchronously from an iothread when handling guest
requests.
Reviewed-by: Pankaj Gupta <pankaj.gupta@ionos.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Message-Id: <20211217134611.31172-7-david@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Let's minimize the number of global variables to prepare for
os_mem_prealloc() getting called concurrently and make the code a bit
easier to read.
The only consumer that really needs a global variable is the sigbus
handler, which will require protection via a mutex in the future either way
as we cannot concurrently mess with the SIGBUS handler.
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Message-Id: <20211217134611.31172-4-david@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Let's sense support and use it for preallocation. MADV_POPULATE_WRITE
does not require a SIGBUS handler, doesn't actually touch page content,
and avoids context switches; it is, therefore, faster and easier to handle
than our current approach.
While MADV_POPULATE_WRITE is, in general, faster than manual
prefaulting, and especially faster with 4k pages, there is still value in
prefaulting using multiple threads to speed up preallocation.
More details on MADV_POPULATE_WRITE can be found in the Linux commits
4ca9b3859dac ("mm/madvise: introduce MADV_POPULATE_(READ|WRITE) to prefault
page tables") and eb2faa513c24 ("mm/madvise: report SIGBUS as -EFAULT for
MADV_POPULATE_(READ|WRITE)"), and in the man page proposal [1].
This resolves the TODO in do_touch_pages().
In the future, we might want to look into using fallocate(), eventually
combined with MADV_POPULATE_READ, when dealing with shared file/fd
mappings and not caring about memory bindings.
[1] https://lkml.kernel.org/r/20210816081922.5155-1-david@redhat.com
Reviewed-by: Pankaj Gupta <pankaj.gupta@ionos.com>
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Message-Id: <20211217134611.31172-3-david@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Let's prepare touch_all_pages() for returning differing errors. Return
an error from the thread and report the last processed error.
Translate SIGBUS to -EFAULT, as a SIGBUS can mean all different kind of
things (memory error, read error, out of memory). When allocating memory
fails via the current SIGBUS-based mechanism, we'll get:
os_mem_prealloc: preallocating memory failed: Bad address
Reviewed-by: Daniel P. Berrangé <berrange@redhat.com>
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Message-Id: <20211217134611.31172-2-david@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
If we warn about the block size being smaller than the default, we skip
some alignment checks.
This can currently only fail on x86-64, when specifying a block size of
1 MiB, however, we detect the THP size of 2 MiB.
Fixes: 228957fea3 ("virtio-mem: Probe THP size to determine default block size")
Cc: "Michael S. Tsirkin" <mst@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Message-Id: <20211011173305.13778-1-david@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
David Hildenbrand