October 2023 - Stratos-dev - op-lists.linaro.org

vhost-device-rng crashes on QEMU > 7.2.0

by Mathieu Poirier

Hey guys, I tried to start the vhost-device-rng daemon with QEMU v8.1.0 when I noticed that upon startup QEMU crashes [1]. The last known good version is v7.2.0. A bisect session between the two tags yielded this commit [2]. In that patch the hunk that starts at line 1176 adds a call to virtio_pci_set_guest_notifier() with a second parameter of VIRTIO_CONFIG_IRQ_IDX, which is -1. That call trickles down to function vhost_user_get_vq_index() [3] where the assert is generated. I am guessing that all the pci based vhost-devices would be affected by this change, hence this email. Has anyone seen this when using a version of QEMU that is higher than v7.2.0? To reproduce, start the vhost-device-rng with: $ ./vhost-device-rng --socket-path=$(PATH_TO_SOCKET)/rng.sock -c 1 -m 512 -p 1000 And in another shell start qemu v8.1.0 with something that looks like this[4], with line 7 to 10 being the most important. Here I'm using RNG but I suppose any of the vhost-devices would be affected by this. I'll wait to hear back from you guys before I make a fool of myself on the QEMU failing list. Thanks, Mathieu [1]. https://pastebin.linaro.org/view/4d39b814 [2]. 1680542862ed virtio-pci: add support for configure interrupt [3]. https://github.com/qemu/qemu/blob/v8.0.0/hw/virtio/vhost-user.c#L2152 [4]. https://pastebin.linaro.org/view/a15db9b0

1 year, 1 month

2
3
0 0

[PATCH V3 1/3] xen-ioctl: Privatize config_irqfd()

by Viresh Kumar

It isn't required to be exposed to client modules, privatize it. Signed-off-by: Viresh Kumar <viresh.kumar(a)linaro.org> --- V3: New commit crates/xen-ioctls/src/xdm/xdm.rs | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/crates/xen-ioctls/src/xdm/xdm.rs b/crates/xen-ioctls/src/xdm/xdm.rs index 3aa5289794fe..06b589967dde 100644 --- a/crates/xen-ioctls/src/xdm/xdm.rs +++ b/crates/xen-ioctls/src/xdm/xdm.rs @@ -242,7 +242,7 @@ impl XenDeviceModelHandle { do_dm_op(&self.fd, domid, &mut privcmd_dm_op_buffers) } - pub fn config_irqfd( + fn config_irqfd( &self, fd: EventFd, domid: u16, -- 2.31.1.272.g89b43f80a514

1 year, 1 month

2
3
0 0

[PATCH 1/2] xen-ioctl: Fix IOCTL macro for irqfd

by Viresh Kumar

The kernel's irqfd macro has changed a bit. Make a corresponding change here too. Signed-off-by: Viresh Kumar <viresh.kumar(a)linaro.org> --- Mathieu, You can apply these patches now. Thanks. crates/xen-ioctls/src/private.rs | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-) diff --git a/crates/xen-ioctls/src/private.rs b/crates/xen-ioctls/src/private.rs index 59a75f67f18d..50833e750f2d 100644 --- a/crates/xen-ioctls/src/private.rs +++ b/crates/xen-ioctls/src/private.rs @@ -12,7 +12,7 @@ use libc::{c_ulong, c_void, ioctl, mmap, munmap, MAP_SHARED, PROT_READ, PROT_WRI use std::fs::OpenOptions; use std::io::Error; use std::os::unix::io::AsRawFd; -use vmm_sys_util::ioctl::_IOC_NONE; +use vmm_sys_util::ioctl::{_IOC_NONE, _IOC_WRITE}; use crate::xdm::types::{PrivcmdDeviceModelIrqFd, PrivcmdDeviceModelOp}; use crate::xfm::types::{PrivCmdMmapBatchV2, PrivCmdMmapResource}; @@ -74,12 +74,11 @@ ioctl_ioc_nr!( ); /* - * #define IOCTL_PRIVCMD_IRQFD \ - * _IOC(_IOC_NONE, 'P', 8, sizeof(privcmd_irqfd_t)) + * #define IOCTL_PRIVCMD_IRQFD _IOW('P', 8, privcmd_irqfd_t) */ ioctl_ioc_nr!( IOCTL_PRIVCMD_IRQFD, - _IOC_NONE, + _IOC_WRITE, XEN_PRIVCMD_TYPE, 8 as u32, std::mem::size_of::<PrivcmdDeviceModelIrqFd>() as u32 -- 2.31.1.272.g89b43f80a514

1 year, 1 month

2
2
0 0

[PATCH V4 0/4] xen: privcmd: Add ioeventfd and fix irqfd support

by Viresh Kumar

Hello, Now that irqfd support (backend to guest interrupt) is already merged, this series solves the other part of the problem, i.e. ioeventfd (guest to backend interrupt). More details inside the commits. Arnd reported few issues with the ioctl macro usage and argument's layout, fixed them for irqfd too, which was added recently. -- Viresh V3->V4: - Use __u64 for indirect pointers in an ioctl command's arguments. - Use u64_to_user_ptr() in kernel driver to access the same. - Use _IOW() macro instead of the internal one: _IOC(). V2->V3: - Remove explicit barriers and depend on spin lock instead to take care of it. - Move check for empty ioeventfds list to privcmd_ioeventfd_deassign(), which could earlier call ioreq_free() even when the list wasn't empty and so we returned without printing a warning in v1 earlier. V2 implemented it incorrectly. V1->V2: - Increment irq_info refcnt only for valid info. - Use u64 type for addr. - Add comments for use of barriers. - Use spin lock instead of mutex as we need to use them in irq handler. - Add a warning when kioreq is getting freed and ioeventfds list isn't empty. - Use struct_size(). - Validate number of vcpus as well. Viresh Kumar (4): xen: Make struct privcmd_irqfd's layout architecture independent xen: irqfd: Use _IOW instead of the internal _IOC() macro xen: evtchn: Allow shared registration of IRQ handers xen: privcmd: Add support for ioeventfd drivers/xen/Kconfig | 8 +- drivers/xen/events/events_base.c | 3 +- drivers/xen/evtchn.c | 2 +- drivers/xen/privcmd.c | 407 +++++++++++++++++++++++++++++- include/uapi/xen/privcmd.h | 22 +- include/xen/interface/hvm/ioreq.h | 51 ++++ 6 files changed, 482 insertions(+), 11 deletions(-) create mode 100644 include/xen/interface/hvm/ioreq.h -- 2.31.1.272.g89b43f80a514

1 year, 1 month

2
8
0 0

[PATCH V3 0/2] xen: privcmd: Add ioeventfd support

by Viresh Kumar

Hello, Now that irqfd support (backend to guest interrupt) is already merged, this series solves the other part of the problem, i.e. ioeventfd (guest to backend interrupt). More details inside the commits. -- Viresh V2->V3: - Remove explicit barriers and depend on spin lock instead to take care of it. - Move check for empty ioeventfds list to privcmd_ioeventfd_deassign(), which could earlier call ioreq_free() even when the list wasn't empty and so we returned without printing a warning in v1 earlier. V2 implemented it incorrectly. V1->V2: - Increment irq_info refcnt only for valid info. - Use u64 type for addr. - Add comments for use of barriers. - Use spin lock instead of mutex as we need to use them in irq handler. - Add a warning when kioreq is getting freed and ioeventfds list isn't empty. - Use struct_size(). - Validate number of vcpus as well. Viresh Kumar (2): xen: evtchn: Allow shared registration of IRQ handers xen: privcmd: Add support for ioeventfd drivers/xen/Kconfig | 8 +- drivers/xen/events/events_base.c | 3 +- drivers/xen/evtchn.c | 2 +- drivers/xen/privcmd.c | 405 +++++++++++++++++++++++++++++- include/uapi/xen/privcmd.h | 18 ++ include/xen/interface/hvm/ioreq.h | 51 ++++ 6 files changed, 479 insertions(+), 8 deletions(-) create mode 100644 include/xen/interface/hvm/ioreq.h -- 2.31.1.272.g89b43f80a514

1 year, 2 months

2
4
0 0

[PATCH 0/2] xen: privcmd: Add ioeventfd support

by Viresh Kumar

Hello, Now that irqfd support (backend to guest interrupt) is already merged, this series solves the other part of the problem, i.e. ioeventfd (guest to backend interrupt). More details inside the commits. -- Viresh Viresh Kumar (2): xen: evtchn: Allow shared registration of IRQ handers xen: privcmd: Add support for ioeventfd drivers/xen/Kconfig | 8 +- drivers/xen/events/events_base.c | 1 + drivers/xen/evtchn.c | 2 +- drivers/xen/privcmd.c | 385 +++++++++++++++++++++++++++++- include/uapi/xen/privcmd.h | 18 ++ include/xen/interface/hvm/ioreq.h | 51 ++++ 6 files changed, 458 insertions(+), 7 deletions(-) create mode 100644 include/xen/interface/hvm/ioreq.h -- 2.31.1.272.g89b43f80a514

1 year, 2 months

4
12
0 0

[PATCH V2 0/2] xen: privcmd: Add ioeventfd support

by Viresh Kumar

Hello, Now that irqfd support (backend to guest interrupt) is already merged, this series solves the other part of the problem, i.e. ioeventfd (guest to backend interrupt). More details inside the commits. -- Viresh V1->V2: - Increment irq_info refcnt only for valid info. - Use u64 type for addr. - Add comments for use of barriers. - Use spin lock instead of mutex as we need to use them in irq handler. - Add a warning when kioreq is getting freed and ioeventfds list isn't empty. - Use struct_size(). - Validate number of vcpus as well. Viresh Kumar (2): xen: evtchn: Allow shared registration of IRQ handers xen: privcmd: Add support for ioeventfd drivers/xen/Kconfig | 8 +- drivers/xen/events/events_base.c | 3 +- drivers/xen/evtchn.c | 2 +- drivers/xen/privcmd.c | 398 +++++++++++++++++++++++++++++- include/uapi/xen/privcmd.h | 18 ++ include/xen/interface/hvm/ioreq.h | 51 ++++ 6 files changed, 472 insertions(+), 8 deletions(-) create mode 100644 include/xen/interface/hvm/ioreq.h -- 2.31.1.272.g89b43f80a514

1 year, 2 months

1
2
0 0

2024

2023

2022

2021

2020

Stratos-dev October 2023