From: Steven Sistare <steven.sistare@oracle.com>
To: "Cédric Le Goater" <clg@redhat.com>, qemu-devel@nongnu.org
Cc: Alex Williamson <alex.williamson@redhat.com>,
Yi Liu <yi.l.liu@intel.com>, Eric Auger <eric.auger@redhat.com>,
Zhenzhong Duan <zhenzhong.duan@intel.com>,
"Michael S. Tsirkin" <mst@redhat.com>,
Marcel Apfelbaum <marcel.apfelbaum@gmail.com>,
Peter Xu <peterx@redhat.com>, Fabiano Rosas <farosas@suse.de>
Subject: Re: [PATCH V1 07/26] vfio/container: recover from unmap-all-vaddr failure
Date: Tue, 4 Feb 2025 11:13:19 -0500 [thread overview]
Message-ID: <61cc8a3c-9af4-466a-b1c1-52692a24b28f@oracle.com> (raw)
In-Reply-To: <a5a4a10a-88d6-478b-b5a9-00f744bf08de@redhat.com>
On 2/4/2025 9:10 AM, Cédric Le Goater wrote:
> On 1/29/25 15:43, Steve Sistare wrote:
>> If there are multiple containers and unmap-all fails for some container, we
>> need to remap vaddr for the other containers for which unmap-all succeeded.
>> Recover by walking all address ranges of all containers to restore the vaddr
>> for each. Do so by invoking the vfio listener callback, and passing a new
>> "remap" flag that tells it to restore a mapping without re-allocating new
>> userland data structures.
>>
>> Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
>> ---
>> hw/vfio/common.c | 47 ++++++++++++++++++++++++++++++++++++++++++-
>> hw/vfio/cpr-legacy.c | 44 ++++++++++++++++++++++++++++++++++++++++
>> include/hw/vfio/vfio-common.h | 6 +++++-
>> 3 files changed, 95 insertions(+), 2 deletions(-)
>>
>> diff --git a/hw/vfio/common.c b/hw/vfio/common.c
>> index 7370332..c8ee71a 100644
>> --- a/hw/vfio/common.c
>> +++ b/hw/vfio/common.c
>> @@ -580,6 +580,13 @@ static void vfio_listener_region_add(MemoryListener *listener,
>> {
>> VFIOContainerBase *bcontainer = container_of(listener, VFIOContainerBase,
>> listener);
>> + vfio_container_region_add(bcontainer, section, false);
>> +}
>> +
>> +void vfio_container_region_add(VFIOContainerBase *bcontainer,
>> + MemoryRegionSection *section,
>> + bool remap)
>> +{
>
> vfio_container_region_add() is already complex enough. Please consider
> doing an initial refactoring before adding a new code path. It would be
> welcome !
I'll take a look after factoring out the cpr code into helpers as you
request below.
>> hwaddr iova, end;
>> Int128 llend, llsize;
>> void *vaddr;
>> @@ -614,6 +621,30 @@ static void vfio_listener_region_add(MemoryListener *listener,
>> int iommu_idx;
>> trace_vfio_listener_region_add_iommu(section->mr->name, iova, end);
>> +
>> + /*
>> + * If remap, then VFIO_DMA_UNMAP_FLAG_VADDR has been called, and we
>> + * want to remap the vaddr. vfio_container_region_add was already
>> + * called in the past, so the giommu already exists. Find it and
>> + * replay it, which calls vfio_dma_map further down the stack.
>> + */
>> +
>> + if (remap) {
>> + hwaddr as_offset = section->offset_within_address_space;
>> + hwaddr iommu_offset = as_offset - section->offset_within_region;
>> +
>> + QLIST_FOREACH(giommu, &bcontainer->giommu_list, giommu_next) {
>> + if (giommu->iommu_mr == iommu_mr &&
>> + giommu->iommu_offset == iommu_offset) {
>> + memory_region_iommu_replay(giommu->iommu_mr, &giommu->n);
>> + return;
>> + }
>> + }
>> + error_report("Container cannot find iommu region %s offset %lx",
>> + memory_region_name(section->mr), iommu_offset);
>
> error_report() are not welcomed. We need to find a way to propagate
> this error.
I follow the existing practice in this function, which already reports other
errors. This is called in the context of a memory listener region_add method,
so returning an error affects the signature of all listeners and should be a
seperate RFE.
>> + goto fail;
>> + }
>
> Please introduce a vfio_cpr helper for the section above and move it
> under the hw/vfio/cpr* files.
OK.
>> /*
>> * FIXME: For VFIO iommu types which have KVM acceleration to
>> * avoid bouncing all map/unmaps through qemu this way, this
>> @@ -656,7 +687,21 @@ static void vfio_listener_region_add(MemoryListener *listener,
>> * about changes.
>> */
>> if (memory_region_has_ram_discard_manager(section->mr)) {
>> - vfio_register_ram_discard_listener(bcontainer, section);
>> + /*
>> + * If remap, then VFIO_DMA_UNMAP_FLAG_VADDR has been called, and we
>> + * want to remap the vaddr. vfio_container_region_add was already
>> + * called in the past, so the ram discard listener already exists.
>> + * Call its populate function directly, which calls vfio_dma_map.
>> + */
>> + if (remap) {
>> + VFIORamDiscardListener *vrdl =
>> + vfio_find_ram_discard_listener(bcontainer, section);
>> + if (vrdl->listener.notify_populate(&vrdl->listener, section)) {
>> + error_report("listener.notify_populate failed");
>> + }
>> + } else {
>> + vfio_register_ram_discard_listener(bcontainer, section);
>> + }
>
> idem.
OK.
>> return;
>> }
>> diff --git a/hw/vfio/cpr-legacy.c b/hw/vfio/cpr-legacy.c
>> index f3a31d1..3139de1 100644
>> --- a/hw/vfio/cpr-legacy.c
>> +++ b/hw/vfio/cpr-legacy.c
>> @@ -26,9 +26,18 @@ static bool vfio_dma_unmap_vaddr_all(VFIOContainer *container, Error **errp)
>> error_setg_errno(errp, errno, "vfio_dma_unmap_vaddr_all");
>> return false;
>> }
>> + container->vaddr_unmapped = true;
>> return true;> }
>> +static void vfio_region_remap(MemoryListener *listener,
>> + MemoryRegionSection *section)
>> +{
>> + VFIOContainer *container = container_of(listener, VFIOContainer,
>> + remap_listener);
>> + vfio_container_region_add(&container->bcontainer, section, true);
>> +}
>> +
>> static bool vfio_cpr_supported(VFIOContainer *container, Error **errp)
>> {
>> if (!ioctl(container->fd, VFIO_CHECK_EXTENSION, VFIO_UPDATE_VADDR)) {
>> @@ -88,6 +97,37 @@ static const VMStateDescription vfio_container_vmstate = {
>> }
>> };
>> +static int vfio_cpr_fail_notifier(NotifierWithReturn *notifier,
>> + MigrationEvent *e, Error **errp)
>> +{
>> + VFIOContainer *container =
>> + container_of(notifier, VFIOContainer, cpr_transfer_notifier);
>> + VFIOContainerBase *bcontainer = &container->bcontainer;
>> +
>> + if (e->type != MIG_EVENT_PRECOPY_FAILED) {
>> + return 0;
>> + }
>> +
>> + if (container->vaddr_unmapped) {
>> + /*
>> + * Force a call to vfio_region_remap for each mapped section by
>> + * temporarily registering a listener, which calls vfio_dma_map
>> + * further down the stack. Set reused so vfio_dma_map restores vaddr.
>> + */
>> + container->reused = true;
>> + container->remap_listener = (MemoryListener) {
>> + .name = "vfio recover",
>> + .region_add = vfio_region_remap
>> + };
>> + memory_listener_register(&container->remap_listener,
>> + bcontainer->space->as);
>> + memory_listener_unregister(&container->remap_listener);
>> + container->reused = false;
>> + container->vaddr_unmapped = false;
>> + }> + return 0;
>> +}
>> +
>> bool vfio_legacy_cpr_register_container(VFIOContainer *container, Error **errp)
>> {
>> VFIOContainerBase *bcontainer = &container->bcontainer;
>> @@ -104,6 +144,9 @@ bool vfio_legacy_cpr_register_container(VFIOContainer *container, Error **errp)
>> vmstate_register(NULL, -1, &vfio_container_vmstate, container);
>> + migration_add_notifier_mode(&container->cpr_transfer_notifier,
>> + vfio_cpr_fail_notifier,
>> + MIG_MODE_CPR_TRANSFER);
>> return true;> }
>> @@ -114,4 +157,5 @@ void vfio_legacy_cpr_unregister_container(VFIOContainer *container)
>> vfio_cpr_unregister_container(bcontainer);
>> migrate_del_blocker(&container->cpr_blocker);
>> vmstate_unregister(NULL, &vfio_container_vmstate, container);
>> + migration_remove_notifier(&container->cpr_transfer_notifier);
>> }
>> diff --git a/include/hw/vfio/vfio-common.h b/include/hw/vfio/vfio-common.h
>> index 1e974e0..8a4a658 100644
>> --- a/include/hw/vfio/vfio-common.h
>> +++ b/include/hw/vfio/vfio-common.h
>> @@ -86,6 +86,9 @@ typedef struct VFIOContainer {
>> unsigned iommu_type;
>> Error *cpr_blocker;
>> bool reused;
>> + bool vaddr_unmapped;
>> + NotifierWithReturn cpr_transfer_notifier;
>> + MemoryListener remap_listener;
>
> There are 5 attributes related to CPR, please add a CPR struct to hold
> them all.
OK.
- Steve
>> QLIST_HEAD(, VFIOGroup) group_list;
>> } VFIOContainer;
>> @@ -311,7 +314,8 @@ int vfio_devices_query_dirty_bitmap(const VFIOContainerBase *bcontainer,
>> int vfio_get_dirty_bitmap(const VFIOContainerBase *bcontainer, uint64_t iova,
>> uint64_t size, ram_addr_t ram_addr, Error **errp);
>> -void vfio_listener_register(VFIOContainerBase *bcontainer);
>> +void vfio_container_region_add(VFIOContainerBase *bcontainer,
>> + MemoryRegionSection *section, bool remap);
>> /* Returns 0 on success, or a negative errno. */
>> bool vfio_device_get_name(VFIODevice *vbasedev, Error **errp);
>
next prev parent reply other threads:[~2025-02-04 16:14 UTC|newest]
Thread overview: 64+ messages / expand[flat|nested] mbox.gz Atom feed top
2025-01-29 14:42 [PATCH V1 00/26] Live update: vfio and iommufd Steve Sistare
2025-01-29 14:42 ` [PATCH V1 01/26] migration: cpr helpers Steve Sistare
2025-01-29 14:42 ` [PATCH V1 02/26] migration: lower handler priority Steve Sistare
2025-02-03 16:21 ` Fabiano Rosas
2025-02-03 16:58 ` Peter Xu
2025-02-06 13:39 ` Steven Sistare
2025-01-29 14:42 ` [PATCH V1 03/26] vfio: vfio_find_ram_discard_listener Steve Sistare
2025-02-03 16:57 ` Cédric Le Goater
2025-01-29 14:43 ` [PATCH V1 04/26] vfio/container: register container for cpr Steve Sistare
2025-02-03 17:01 ` Cédric Le Goater
2025-02-03 22:26 ` Steven Sistare
2025-01-29 14:43 ` [PATCH V1 05/26] vfio/container: preserve descriptors Steve Sistare
2025-02-03 17:48 ` Cédric Le Goater
2025-02-03 22:26 ` Steven Sistare
2025-01-29 14:43 ` [PATCH V1 06/26] vfio/container: preserve DMA mappings Steve Sistare
2025-02-03 18:25 ` Cédric Le Goater
2025-02-03 22:27 ` Steven Sistare
2025-01-29 14:43 ` [PATCH V1 07/26] vfio/container: recover from unmap-all-vaddr failure Steve Sistare
2025-02-04 14:10 ` Cédric Le Goater
2025-02-04 16:13 ` Steven Sistare [this message]
2025-01-29 14:43 ` [PATCH V1 08/26] pci: skip reset during cpr Steve Sistare
2025-02-04 14:14 ` Cédric Le Goater
2025-02-04 16:13 ` Steven Sistare
2025-01-29 14:43 ` [PATCH V1 09/26] pci: export msix_is_pending Steve Sistare
2025-01-29 14:43 ` [PATCH V1 10/26] vfio-pci: refactor for cpr Steve Sistare
2025-02-04 14:39 ` Cédric Le Goater
2025-02-04 16:14 ` Steven Sistare
2025-01-29 14:43 ` [PATCH V1 11/26] vfio-pci: skip reset during cpr Steve Sistare
2025-02-04 14:56 ` Cédric Le Goater
2025-02-04 16:15 ` Steven Sistare
2025-01-29 14:43 ` [PATCH V1 12/26] vfio-pci: preserve MSI Steve Sistare
2025-02-05 16:48 ` Cédric Le Goater
2025-02-06 14:41 ` Steven Sistare
2025-01-29 14:43 ` [PATCH V1 13/26] vfio-pci: preserve INTx Steve Sistare
2025-02-05 17:13 ` Cédric Le Goater
2025-02-06 14:43 ` Steven Sistare
2025-01-29 14:43 ` [PATCH V1 14/26] migration: close kvm after cpr Steve Sistare
2025-01-29 14:43 ` [PATCH V1 15/26] migration: cpr_get_fd_param helper Steve Sistare
2025-01-29 14:43 ` [PATCH V1 16/26] vfio: return mr from vfio_get_xlat_addr Steve Sistare
2025-02-04 15:47 ` Cédric Le Goater
2025-02-04 17:42 ` Steven Sistare
2025-02-16 23:19 ` John Levon
2025-01-29 14:43 ` [PATCH V1 17/26] vfio: pass ramblock to vfio_container_dma_map Steve Sistare
2025-01-29 14:43 ` [PATCH V1 18/26] vfio/iommufd: define iommufd_cdev_make_hwpt Steve Sistare
2025-02-04 16:22 ` Cédric Le Goater
2025-02-04 17:42 ` Steven Sistare
2025-01-29 14:43 ` [PATCH V1 19/26] vfio/iommufd: use IOMMU_IOAS_MAP_FILE Steve Sistare
2025-02-05 17:23 ` Cédric Le Goater
2025-02-05 22:01 ` Steven Sistare
2025-01-29 14:43 ` [PATCH V1 20/26] vfio/iommufd: export iommufd_cdev_get_info_iova_range Steve Sistare
2025-02-05 17:33 ` Cédric Le Goater
2025-02-05 22:01 ` Steven Sistare
2025-01-29 14:43 ` [PATCH V1 21/26] iommufd: change process ioctl Steve Sistare
2025-02-05 17:34 ` Cédric Le Goater
2025-02-05 22:02 ` Steven Sistare
2025-01-29 14:43 ` [PATCH V1 22/26] vfio/iommufd: invariant device name Steve Sistare
2025-02-05 17:42 ` Cédric Le Goater
2025-02-05 22:02 ` Steven Sistare
2025-01-29 14:43 ` [PATCH V1 23/26] vfio/iommufd: register container for cpr Steve Sistare
2025-02-05 17:45 ` Cédric Le Goater
2025-02-05 22:03 ` Steven Sistare
2025-01-29 14:43 ` [PATCH V1 24/26] vfio/iommufd: preserve descriptors Steve Sistare
2025-01-29 14:43 ` [PATCH V1 25/26] vfio/iommufd: reconstruct device Steve Sistare
2025-01-29 14:43 ` [PATCH V1 26/26] iommufd: preserve DMA mappings Steve Sistare
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=61cc8a3c-9af4-466a-b1c1-52692a24b28f@oracle.com \
--to=steven.sistare@oracle.com \
--cc=alex.williamson@redhat.com \
--cc=clg@redhat.com \
--cc=eric.auger@redhat.com \
--cc=farosas@suse.de \
--cc=marcel.apfelbaum@gmail.com \
--cc=mst@redhat.com \
--cc=peterx@redhat.com \
--cc=qemu-devel@nongnu.org \
--cc=yi.l.liu@intel.com \
--cc=zhenzhong.duan@intel.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).