lkml.org 
[lkml]   [2022]   [May]   [30]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
SubjectRe: Re: [PATCH 3/3] virtio_balloon: Introduce memory recover
From


On 5/30/22 15:48, David Hildenbrand wrote:
>
>> +
>> struct virtio_balloon {
>> struct virtio_device *vdev;
>> struct virtqueue *inflate_vq, *deflate_vq, *stats_vq, *free_page_vq;
>> @@ -126,6 +133,16 @@ struct virtio_balloon {
>> /* Free page reporting device */
>> struct virtqueue *reporting_vq;
>> struct page_reporting_dev_info pr_dev_info;
>> +
>> + /* Memory recover VQ - VIRTIO_BALLOON_F_RECOVER */
>> + struct virtqueue *recover_vq;
>> + spinlock_t recover_vq_lock;
>> + struct notifier_block memory_failure_nb;
>> + struct list_head corrupted_page_list;
>> + struct list_head recovered_page_list;
>> + spinlock_t recover_page_list_lock;
>> + struct __virtio_balloon_recover in_vbr;
>> + struct work_struct unpoison_memory_work;
>
> I assume we want all that only with CONFIG_MEMORY_FAILURE.
>

Sorry, I missed this.

>> };
>>
>> static const struct virtio_device_id id_table[] = {
>> @@ -494,6 +511,198 @@ static void update_balloon_size_func(struct work_struct *work)
>> queue_work(system_freezable_wq, work);
>> }
>>
>> +/*
>> + * virtballoon_memory_failure - notified by memory failure, try to fix the
>> + * corrupted page.
>> + * The memory failure notifier is designed to call back when the kernel handled
>> + * successfully only, WARN_ON_ONCE on the unlikely condition to find out any
>> + * error(memory error handling is a best effort, not 100% coverd).
>> + */
>> +static int virtballoon_memory_failure(struct notifier_block *notifier,
>> + unsigned long pfn, void *parm)
>> +{
>> + struct virtio_balloon *vb = container_of(notifier, struct virtio_balloon,
>> + memory_failure_nb);
>> + struct page *page;
>> + struct __virtio_balloon_recover *out_vbr;
>> + struct scatterlist sg;
>> + unsigned long flags;
>> + int err;
>> +
>> + page = pfn_to_online_page(pfn);
>> + if (WARN_ON_ONCE(!page))
>> + return NOTIFY_DONE;
>> +
>> + if (PageHuge(page))
>> + return NOTIFY_DONE;
>> +
>> + if (WARN_ON_ONCE(!PageHWPoison(page)))
>> + return NOTIFY_DONE;
>> +
>> + if (WARN_ON_ONCE(page_count(page) != 1))
>> + return NOTIFY_DONE;
>
> Relying on the page_count to be 1 for correctness is usually a bit
> shaky, for example, when racing against isolate_movable_page() that
> might temporarily bump upo the refcount.
>

The memory notifier is designed to call the chain if a page gets result
MF_RECOVERED only:
if (result == MF_RECOVERED)
blocking_notifier_call_chain(&mf_notifier_list, pfn, NULL);


>> +
>> + get_page(page); /* balloon reference */
>> +
>> + out_vbr = kzalloc(sizeof(*out_vbr), GFP_KERNEL);
>
> Are we always guaranteed to be able to use GFP_KERNEL out of MCE
> context? (IOW, are we never atomic?)
>
>> + if (WARN_ON_ONCE(!out_vbr))
>> + return NOTIFY_BAD;
>> +
>> + spin_lock(&vb->recover_page_list_lock);
>> + balloon_page_push(&vb->corrupted_page_list, page);
>> + spin_unlock(&vb->recover_page_list_lock);
>> +
>> + out_vbr->vbr.cmd = VIRTIO_BALLOON_R_CMD_RECOVER;
>
> This makes me wonder if we should have a more generic guest->host
> request queue, similar to what e.g., virtio-mem uses, instead of adding
> a separate VIRTIO_BALLOON_VQ_RECOVER vq.
>

I'm OK with either one, I'll follow your decision! :D

>> + set_page_pfns(vb, out_vbr->pfns, page);
>> + sg_init_one(&sg, out_vbr, sizeof(*out_vbr));
>> +
>> + spin_lock_irqsave(&vb->recover_vq_lock, flags);
>> + err = virtqueue_add_outbuf(vb->recover_vq, &sg, 1, out_vbr, GFP_KERNEL);
>> + if (unlikely(err)) {
>> + spin_unlock_irqrestore(&vb->recover_vq_lock, flags);
>> + return NOTIFY_DONE;
>> + }
>> + virtqueue_kick(vb->recover_vq);
>> + spin_unlock_irqrestore(&vb->recover_vq_lock, flags);
>> +
>> + return NOTIFY_OK;
>> +}
>> +
>> +static int recover_vq_get_response(struct virtio_balloon *vb)
>> +{
>> + struct __virtio_balloon_recover *in_vbr;
>> + struct scatterlist sg;
>> + unsigned long flags;
>> + int err;
>> +
>> + spin_lock_irqsave(&vb->recover_vq_lock, flags);
>> + in_vbr = &vb->in_vbr;
>> + memset(in_vbr, 0x00, sizeof(*in_vbr));
>> + sg_init_one(&sg, in_vbr, sizeof(*in_vbr));
>> + err = virtqueue_add_inbuf(vb->recover_vq, &sg, 1, in_vbr, GFP_KERNEL);
>> + if (unlikely(err)) {
>> + spin_unlock_irqrestore(&vb->recover_vq_lock, flags);
>> + return err;
>> + }
>> +
>> + virtqueue_kick(vb->recover_vq);
>> + spin_unlock_irqrestore(&vb->recover_vq_lock, flags);
>> +
>> + return 0;
>> +}
>> +
>> +static void recover_vq_handle_response(struct virtio_balloon *vb, unsigned int len)
>> +{
>> + struct __virtio_balloon_recover *in_vbr;
>> + struct virtio_balloon_recover *vbr;
>> + struct page *page;
>> + unsigned int pfns;
>> + u32 pfn0, pfn1;
>> + __u8 status;
>> +
>> + /* the response is not expected */
>> + if (unlikely(len != sizeof(struct __virtio_balloon_recover)))
>> + return;
>> +
>> + in_vbr = &vb->in_vbr;
>> + vbr = &in_vbr->vbr;
>> + if (unlikely(vbr->cmd != VIRTIO_BALLOON_R_CMD_RESPONSE))
>> + return;
>> +
>> + /* to make sure the contiguous balloon PFNs */
>> + for (pfns = 1; pfns < VIRTIO_BALLOON_PAGES_PER_PAGE; pfns++) {
>> + pfn0 = virtio32_to_cpu(vb->vdev, in_vbr->pfns[pfns - 1]);
>> + pfn1 = virtio32_to_cpu(vb->vdev, in_vbr->pfns[pfns]);
>> + if (pfn1 - pfn0 != 1)
>> + return;
>
> Yeah, we really shouldn't be dealing with (legacy) 4k PFNs here, but
> instead, proper ranges I guess.
>

MST also pointed out this, I explained in this link:
https://lkml.org/lkml/2022/5/26/942

Rather than page reporting style, virtio-mem style should be fine. Ex,
struct virtio_memory_recover {
__virtio64 addr;
__virtio32 length;
__virtio16 padding[2];
};

>> + }
>> +
>> + pfn0 = virtio32_to_cpu(vb->vdev, in_vbr->pfns[0]);
>> + if (!pfn_valid(pfn0))
>> + return;
>> +
>> + pfn1 = -1;
>> + spin_lock(&vb->recover_page_list_lock);
>> + list_for_each_entry(page, &vb->corrupted_page_list, lru) {
>> + pfn1 = page_to_pfn(page);
>> + if (pfn1 == pfn0)
>> + break;
>> + }
>> + spin_unlock(&vb->recover_page_list_lock);
>> +
>> + status = vbr->status;
>> + switch (status) {
>> + case VIRTIO_BALLOON_R_STATUS_RECOVERED:
>> + if (pfn1 == pfn0) {
>> + spin_lock(&vb->recover_page_list_lock);
>> + list_del(&page->lru);
>> + balloon_page_push(&vb->recovered_page_list, page);
>
> We rather not reuse actual balloon functions in !balloon context. Just
> move the page to the proper list directly.
>

OK.

>> + spin_unlock(&vb->recover_page_list_lock);
>> + queue_work(system_freezable_wq, &vb->unpoison_memory_work);
>> + dev_info_ratelimited(&vb->vdev->dev, "recovered pfn 0x%x", pfn0);
>
> Well, not yet. Shouldn't this go into unpoison_memory_func() ?
>

OK.

[...]

>
>>
>> +out_unregister_reporting:
>> + if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_REPORTING))
>> + page_reporting_unregister(&vb->pr_dev_info);
>> out_unregister_oom:
>> if (virtio_has_feature(vb->vdev, VIRTIO_BALLOON_F_DEFLATE_ON_OOM))
>> unregister_oom_notifier(&vb->oom_nb);
>> @@ -1082,6 +1319,11 @@ static void virtballoon_remove(struct virtio_device *vdev)
>> destroy_workqueue(vb->balloon_wq);
>> }
>>
>> + if (virtio_has_feature(vdev, VIRTIO_BALLOON_F_RECOVER)) {
>
> Could the notifier already have been triggered and we might be using the
> device before already fully initialized from the notifier and might end
> up leaking memory here that we allocated?
>
>> + unregister_memory_failure_notifier(&vb->memory_failure_nb);
>> + cancel_work_sync(&vb->unpoison_memory_work);
>> + }
>> +
>
> Could we be leaking memory from the virtballoon_remove() path?
>

Yes, I'll fix the possible memory leak here.

Thanks a lot.

--
zhenwei pi

\
 
 \ /
  Last update: 2022-05-30 14:52    [W:0.180 / U:0.664 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site