We already use MADV_NORESERVE to deal with sparse memory regions. Let's
also set madvise(MADV_DONTDUMP), otherwise a crash of the process can
result in us allocating all memory in the mmap'ed region for dumping
purposes.
This change implies that the mmap'ed rings won't be included in a
coredump. If ever required for debugging purposes, we could mark only
the mapped rings MADV_DODUMP.
Ignore errors during madvise() for now.
Reviewed-by: Raphael Norwitz <raphael@enfabrica.net>
Acked-by: Stefano Garzarella <sgarzare@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Message-Id: <20240214151701.29906-15-david@redhat.com>
Tested-by: Mario Casquero <mcasquer@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Currently, we try to remap all rings whenever we add a single new memory
region. That doesn't quite make sense, because we already map rings when
setting the ring address, and panic if that goes wrong. Likely, that
handling was simply copied from set_mem_table code, where we actually
have to remap all rings.
Remapping all rings might require us to walk quite a lot of memory
regions to perform the address translations. Ideally, we'd simply remove
that remapping.
However, let's be a bit careful. There might be some weird corner cases
where we might temporarily remove a single memory region (e.g., resize
it), that would have worked for now. Further, a ring might be located on
hotplugged memory, and as the VM reboots, we might unplug that memory, to
hotplug memory before resetting the ring addresses.
So let's unmap affected rings as we remove a memory region, and try
dynamically mapping the ring again when required.
Acked-by: Raphael Norwitz <raphael@enfabrica.net>
Acked-by: Stefano Garzarella <sgarzare@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Message-Id: <20240214151701.29906-14-david@redhat.com>
Tested-by: Mario Casquero <mcasquer@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
In the past, QEMU would create memory regions that could partially cover
hugetlb pages, making mmap() fail if we would use the mmap_offset as an
fd_offset. For that reason, we never used the mmap_offset as an offset into
the fd and instead always mapped the fd from the very start.
However, that can easily result in us mmap'ing a lot of unnecessary
parts of an fd, possibly repeatedly.
QEMU nowadays does not create memory regions that partially cover huge
pages -- it never really worked with postcopy. QEMU handles merging of
regions that partially cover huge pages (due to holes in boot memory) since
2018 in c1ece84e7c ("vhost: Huge page align and merge").
Let's be a bit careful and not unconditionally convert the
mmap_offset into an fd_offset. Instead, let's simply detect the hugetlb
size and pass as much as we can as fd_offset, making sure that we call
mmap() with a properly aligned offset.
With QEMU and a virtio-mem device that is fully plugged (50GiB using 50
memslots) the qemu-storage daemon process consumes in the VA space
1281GiB before this change and 58GiB after this change.
================ Vhost user message ================
Request: VHOST_USER_ADD_MEM_REG (37)
Flags: 0x9
Size: 40
Fds: 59
Adding region 4
guest_phys_addr: 0x0000000200000000
memory_size: 0x0000000040000000
userspace_addr: 0x00007fb73bffe000
old mmap_offset: 0x0000000080000000
fd_offset: 0x0000000080000000
new mmap_offset: 0x0000000000000000
mmap_addr: 0x00007f02f1bdc000
Successfully added new region
================ Vhost user message ================
Request: VHOST_USER_ADD_MEM_REG (37)
Flags: 0x9
Size: 40
Fds: 59
Adding region 5
guest_phys_addr: 0x0000000240000000
memory_size: 0x0000000040000000
userspace_addr: 0x00007fb77bffe000
old mmap_offset: 0x00000000c0000000
fd_offset: 0x00000000c0000000
new mmap_offset: 0x0000000000000000
mmap_addr: 0x00007f0284000000
Successfully added new region
Reviewed-by: Raphael Norwitz <raphael@enfabrica.net>
Acked-by: Stefano Garzarella <sgarzare@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Message-Id: <20240214151701.29906-12-david@redhat.com>
Tested-by: Mario Casquero <mcasquer@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Let's speed up GPA to memory region / virtual address lookup. Store the
memory regions ordered by guest physical addresses, and use binary
search for address translation, as well as when adding/removing memory
regions.
Most importantly, this will speed up GPA->VA address translation when we
have many memslots.
Reviewed-by: Raphael Norwitz <raphael@enfabrica.net>
Acked-by: Stefano Garzarella <sgarzare@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Message-Id: <20240214151701.29906-11-david@redhat.com>
Tested-by: Mario Casquero <mcasquer@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Memory regions cannot overlap, and if we ever hit that case something
would be really flawed.
For example, when vhost code in QEMU decides to increase the size of memory
regions to cover full huge pages, it makes sure to never create overlaps,
and if there would be overlaps, it would bail out.
QEMU commits 48d7c97577 ("vhost: Merge sections added to temporary
list"), c1ece84e7c ("vhost: Huge page align and merge") and
e7b94a84b6 ("vhost: Allow adjoining regions") added and clarified that
handling and how overlaps are impossible.
Consequently, each GPA can belong to at most one memory region, and
everything else doesn't make sense. Let's factor out our search to prepare
for further changes.
Reviewed-by: Raphael Norwitz <raphael@enfabrica.net>
Acked-by: Stefano Garzarella <sgarzare@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Message-Id: <20240214151701.29906-10-david@redhat.com>
Tested-by: Mario Casquero <mcasquer@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Let's factor it out, reducing quite some code duplication and perparing
for further changes.
If we fail to mmap a region and panic, we now simply don't add that
(broken) region.
Note that we now increment dev->nregions as we are successfully
adding memory regions, and don't increment dev->nregions if anything went
wrong.
Reviewed-by: Raphael Norwitz <raphael@enfabrica.net>
Acked-by: Stefano Garzarella <sgarzare@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Message-Id: <20240214151701.29906-6-david@redhat.com>
Tested-by: Mario Casquero <mcasquer@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Let's support up to 509 mem slots, just like vhost in the kernel usually
does and the rust vhost-user implementation recently [1] started doing.
This is required to properly support memory hotplug, either using
multiple DIMMs (ACPI supports up to 256) or using virtio-mem.
The 509 used to be the KVM limit, it supported 512, but 3 were
used for internal purposes. Currently, KVM supports more than 512, but
it usually doesn't make use of more than ~260 (i.e., 256 DIMMs + boot
memory), except when other memory devices like PCI devices with BARs are
used. So, 509 seems to work well for vhost in the kernel.
Details can be found in the QEMU change that made virtio-mem consume
up to 256 mem slots across all virtio-mem devices. [2]
509 mem slots implies 509 VMAs/mappings in the worst case (even though,
in practice with virtio-mem we won't be seeing more than ~260 in most
setups).
With max_map_count under Linux defaulting to 64k, 509 mem slots
still correspond to less than 1% of the maximum number of mappings.
There are plenty left for the application to consume.
[1] https://github.com/rust-vmm/vhost/pull/224
[2] https://lore.kernel.org/all/20230926185738.277351-1-david@redhat.com/
Reviewed-by: Raphael Norwitz <raphael@enfabrica.net>
Acked-by: Stefano Garzarella <sgarzare@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Message-Id: <20240214151701.29906-3-david@redhat.com>
Tested-by: Mario Casquero <mcasquer@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Let's prepare for increasing VHOST_USER_MAX_RAM_SLOTS by dynamically
allocating dev->regions. We don't have any ABI guarantees (not
dynamically linked), so we can simply change the layout of VuDev.
Let's zero out the memory, just as we used to do.
Reviewed-by: Raphael Norwitz <raphael@enfabrica.net>
Acked-by: Stefano Garzarella <sgarzare@redhat.com>
Signed-off-by: David Hildenbrand <david@redhat.com>
Message-Id: <20240214151701.29906-2-david@redhat.com>
Tested-by: Mario Casquero <mcasquer@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
The VHOST_USER_PROTOCOL_F_XEN_MMAP feature bit was defined in
f21e95ee97, which has been part of qemu's 8.1.0 release. However, it
seems it was never added to qemu's code, but it is well possible that it
is already used by different front-ends outside of qemu (i.e., Xen).
VHOST_USER_PROTOCOL_F_SHARED_OBJECT in contrast was added to qemu's code
in 1609476662, but never defined in the vhost-user specification. As a
consequence, both bits were defined to be 17, which cannot work.
Regardless of whether actual code or the specification should take
precedence, F_XEN_MMAP is already part of a qemu release, while
F_SHARED_OBJECT is not. Therefore, bump the latter to take number 18
instead of 17, and add this to the specification.
Take the opportunity to add at least a little note on the
VhostUserShared structure to the specification. This structure is
referenced by the new commands introduced in 1609476662, but was not
defined.
Fixes: 1609476662
("vhost-user: add shared_object msg")
Signed-off-by: Hanna Czenczek <hreitz@redhat.com>
Message-Id: <20231016083201.23736-1-hreitz@redhat.com>
Reviewed-by: Emmanouil Pitsidianakis <manos.pitsidianakis@linaro.org>
Reviewed-by: Stefano Garzarella <sgarzare@redhat.com>
Reviewed-by: Viresh Kumar <viresh.kumar@linaro.org>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
In the libvhost-user library we need to
handle VHOST_USER_GET_SHARED_OBJECT requests,
and add helper functions to allow sending messages
to interact with the virtio shared objects
hash table.
Signed-off-by: Albert Esteve <aesteve@redhat.com>
Message-Id: <20231002065706.94707-5-aesteve@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
The source file uses VIRTIO_F_VERSION_1 which is
not defined by <linux/virtio_config.h> on Debian 10.
The system-provided <linux/virtio_config.h> which
does not include the macro definition is included
through <linux/vhost.h>, so fix the issue by including
the standard-headers version before that.
Signed-off-by: David 'Digit' Turner <digit@google.com>
Message-Id: <20230405125920.2951721-2-digit@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Currently block_resize qmp command is simply ignored by vhost-user-blk
export. So, the block-node is successfully resized, but virtio config
is unchanged and guest doesn't see that disk is resized.
Let's handle the resize by modifying the config and notifying the guest
appropriately.
After this comment, lsblk in linux guest with attached
vhost-user-blk-pci device shows new size immediately after block_resize
QMP command on vhost-user exported block node.
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@yandex-team.ru>
Message-Id: <20230321201323.3695923-1-vsementsov@yandex-team.ru>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
