On 01/03/2019 12:38, Alexey Kardashevskiy wrote:
> We already allocate hardware TCE tables in multiple levels and skip
> intermediate levels when we can, now it is a turn of the KVM TCE tables.
> Thankfully these are allocated already in 2 levels.
>
> This moves the table's last level allocation from the creating helper to
> kvmppc_tce_put() and kvm_spapr_tce_fault().
>
> This adds kvmppc_rm_ioba_validate() to do an additional test if
> the consequent kvmppc_tce_put() needs a page which has not been allocated;
> if this is the case, we bail out to virtual mode handlers.
>
> Signed-off-by: Alexey Kardashevskiy <a...@ozlabs.ru>
> ---
> Changes:
> v2:
> * added kvm mutex around alloc_page to prevent races; in both place we
> test the pointer, if NULL, then take a lock and check again so on a fast
> path we do not take a lock at all
>
>
> ---
> For NVLink2 passthrough guests with 128TiB DMA windows and very fragmented
> system RAM the difference is gigabytes of RAM.
> ---
> arch/powerpc/kvm/book3s_64_vio.c | 29 ++++++------
> arch/powerpc/kvm/book3s_64_vio_hv.c | 69 ++++++++++++++++++++++++++---
> 2 files changed, 79 insertions(+), 19 deletions(-)
>
> diff --git a/arch/powerpc/kvm/book3s_64_vio.c
> b/arch/powerpc/kvm/book3s_64_vio.c
> index f02b049..7eed8c9 100644
> --- a/arch/powerpc/kvm/book3s_64_vio.c
> +++ b/arch/powerpc/kvm/book3s_64_vio.c
> @@ -228,7 +228,8 @@ static void release_spapr_tce_table(struct rcu_head *head)
> unsigned long i, npages = kvmppc_tce_pages(stt->size);
>
> for (i = 0; i < npages; i++)
> - __free_page(stt->pages[i]);
> + if (stt->pages[i])
> + __free_page(stt->pages[i]);
>
> kfree(stt);
> }
> @@ -242,6 +243,20 @@ static vm_fault_t kvm_spapr_tce_fault(struct vm_fault
> *vmf)
> return VM_FAULT_SIGBUS;
>
> page = stt->pages[vmf->pgoff];
> + if (!page) {
> + mutex_lock(&stt->kvm->lock);
> + page = stt->pages[vmf->pgoff];
> + if (!page) {
> + page = alloc_page(GFP_KERNEL | __GFP_ZERO);
> + if (!page) {
> + mutex_unlock(&stt->kvm->lock);
> + return VM_FAULT_OOM;
> + }
> + stt->pages[vmf->pgoff] = page;
> + }
> + mutex_unlock(&stt->kvm->lock);
> + }
> +
> get_page(page);
> vmf->page = page;
> return 0;
> @@ -296,7 +311,6 @@ long kvm_vm_ioctl_create_spapr_tce(struct kvm *kvm,
> struct kvmppc_spapr_tce_table *siter;
> unsigned long npages, size = args->size;
> int ret = -ENOMEM;
> - int i;
>
> if (!args->size || args->page_shift < 12 || args->page_shift > 34 ||
> (args->offset + args->size > (ULLONG_MAX >> args->page_shift)))
> @@ -320,12 +334,6 @@ long kvm_vm_ioctl_create_spapr_tce(struct kvm *kvm,
> stt->kvm = kvm;
> INIT_LIST_HEAD_RCU(&stt->iommu_tables);
>
> - for (i = 0; i < npages; i++) {
> - stt->pages[i] = alloc_page(GFP_KERNEL | __GFP_ZERO);
> - if (!stt->pages[i])
> - goto fail;
> - }
> -
> mutex_lock(&kvm->lock);
>
> /* Check this LIOBN hasn't been previously allocated */
> @@ -352,11 +360,6 @@ long kvm_vm_ioctl_create_spapr_tce(struct kvm *kvm,
> if (ret >= 0)
> return ret;
>
> - fail:
> - for (i = 0; i < npages; i++)
> - if (stt->pages[i])
> - __free_page(stt->pages[i]);
> -
> kfree(stt);
> fail_acct:
> kvmppc_account_memlimit(kvmppc_stt_pages(npages), false);
> diff --git a/arch/powerpc/kvm/book3s_64_vio_hv.c
> b/arch/powerpc/kvm/book3s_64_vio_hv.c
> index 2206bc7..a0912d5 100644
> --- a/arch/powerpc/kvm/book3s_64_vio_hv.c
> +++ b/arch/powerpc/kvm/book3s_64_vio_hv.c
> @@ -158,23 +158,76 @@ static u64 *kvmppc_page_address(struct page *page)
> return (u64 *) page_address(page);
> }
>
> +/*
> + * TCEs pages are allocated in kvmppc_tce_put() which won't be able to do so
> + * in real mode.
> + * Check if kvmppc_tce_put() can succeed in real mode, i.e. a TCEs page is
> + * allocated or not required (when clearing a tce entry).
> + */
> +static long kvmppc_rm_ioba_validate(struct kvmppc_spapr_tce_table *stt,
> + unsigned long ioba, unsigned long npages, bool clearing)
> +{
> + unsigned long i, sttpage, sttpages;
> + unsigned long ret = kvmppc_ioba_validate(stt, ioba, npages);
> +
> + if (ret)
> + return ret;
> + /*
> + * clearing==true says kvmppc_tce_put won't be allocating pages
> + * for empty tces.
> + */
> + if (clearing)
> + return H_SUCCESS;
> +
> + sttpage = ((ioba >> stt->page_shift) - stt->offset) / TCES_PER_PAGE;
> + sttpages = (npages + TCES_PER_PAGE - 1) / TCES_PER_PAGE;
This is wrong, v3 is coming.
--
Alexey
