Improve make_secure_pte to avoid stalls when the system is heavily
overcommitted. This was especially problematic in kvm_s390_pv_unpack,
because of the loop over all pages that needed unpacking.
Due to the locks being held, it was not possible to simply replace
uv_call with uv_call_sched. A more complex approach was
needed, in which uv_call is replaced with __uv_call, which does not
loop. When the UVC needs to be executed again, -EAGAIN is returned, and
the caller (or its caller) will try again.
When -EAGAIN is returned, the path is the same as when the page is in
writeback (and the writeback check is also performed, which is
harmless).
Signed-off-by: Claudio Imbrenda <[email protected]>
Fixes: 214d9bbcd3a672 ("s390/mm: provide memory management functions for protected KVM guests")
---
arch/s390/kernel/uv.c | 29 +++++++++++++++++++++++------
arch/s390/kvm/intercept.c | 5 +++++
2 files changed, 28 insertions(+), 6 deletions(-)
diff --git a/arch/s390/kernel/uv.c b/arch/s390/kernel/uv.c
index aeb0a15bcbb7..68a8fbafcb9c 100644
--- a/arch/s390/kernel/uv.c
+++ b/arch/s390/kernel/uv.c
@@ -180,7 +180,7 @@ static int make_secure_pte(pte_t *ptep, unsigned long addr,
{
pte_t entry = READ_ONCE(*ptep);
struct page *page;
- int expected, rc = 0;
+ int expected, cc = 0;
if (!pte_present(entry))
return -ENXIO;
@@ -196,12 +196,25 @@ static int make_secure_pte(pte_t *ptep, unsigned long addr,
if (!page_ref_freeze(page, expected))
return -EBUSY;
set_bit(PG_arch_1, &page->flags);
- rc = uv_call(0, (u64)uvcb);
+ /*
+ * If the UVC does not succeed or fail immediately, we don't want to
+ * loop for long, or we might get stall notifications.
+ * On the other hand, this is a complex scenario and we are holding a lot of
+ * locks, so we can't easily sleep and reschedule. We try only once,
+ * and if the UVC returned busy or partial completion, we return
+ * -EAGAIN and we let the callers deal with it.
+ */
+ cc = __uv_call(0, (u64)uvcb);
page_ref_unfreeze(page, expected);
- /* Return -ENXIO if the page was not mapped, -EINVAL otherwise */
- if (rc)
- rc = uvcb->rc == 0x10a ? -ENXIO : -EINVAL;
- return rc;
+ /*
+ * Return -ENXIO if the page was not mapped, -EINVAL for other errors.
+ * If busy or partially completed, return -EAGAIN.
+ */
+ if (cc == UVC_CC_OK)
+ return 0;
+ else if (cc == UVC_CC_BUSY || cc == UVC_CC_PARTIAL)
+ return -EAGAIN;
+ return uvcb->rc == 0x10a ? -ENXIO : -EINVAL;
}
/*
@@ -254,6 +267,10 @@ int gmap_make_secure(struct gmap *gmap, unsigned long gaddr, void *uvcb)
mmap_read_unlock(gmap->mm);
if (rc == -EAGAIN) {
+ /*
+ * If we are here because the UVC returned busy or partial
+ * completion, this is just a useless check, but it is safe.
+ */
wait_on_page_writeback(page);
} else if (rc == -EBUSY) {
/*
diff --git a/arch/s390/kvm/intercept.c b/arch/s390/kvm/intercept.c
index 72b25b7cc6ae..47833ade4da5 100644
--- a/arch/s390/kvm/intercept.c
+++ b/arch/s390/kvm/intercept.c
@@ -516,6 +516,11 @@ static int handle_pv_uvc(struct kvm_vcpu *vcpu)
*/
if (rc == -EINVAL)
return 0;
+ /*
+ * If we got -EAGAIN here, we simply return it. It will eventually
+ * get propagated all the way to userspace, which should then try
+ * again.
+ */
return rc;
}
--
2.31.1
Am 20.09.21 um 15:24 schrieb Claudio Imbrenda:
> Improve make_secure_pte to avoid stalls when the system is heavily
> overcommitted. This was especially problematic in kvm_s390_pv_unpack,
> because of the loop over all pages that needed unpacking.
>
> Due to the locks being held, it was not possible to simply replace
> uv_call with uv_call_sched. A more complex approach was
> needed, in which uv_call is replaced with __uv_call, which does not
> loop. When the UVC needs to be executed again, -EAGAIN is returned, and
> the caller (or its caller) will try again.
>
> When -EAGAIN is returned, the path is the same as when the page is in
> writeback (and the writeback check is also performed, which is
> harmless).
>
> Signed-off-by: Claudio Imbrenda <[email protected]>
> Fixes: 214d9bbcd3a672 ("s390/mm: provide memory management functions for protected KVM guests")
> ---
> arch/s390/kernel/uv.c | 29 +++++++++++++++++++++++------
> arch/s390/kvm/intercept.c | 5 +++++
> 2 files changed, 28 insertions(+), 6 deletions(-)
>
> diff --git a/arch/s390/kernel/uv.c b/arch/s390/kernel/uv.c
> index aeb0a15bcbb7..68a8fbafcb9c 100644
> --- a/arch/s390/kernel/uv.c
> +++ b/arch/s390/kernel/uv.c
> @@ -180,7 +180,7 @@ static int make_secure_pte(pte_t *ptep, unsigned long addr,
> {
> pte_t entry = READ_ONCE(*ptep);
> struct page *page;
> - int expected, rc = 0;
> + int expected, cc = 0;
>
> if (!pte_present(entry))
> return -ENXIO;
> @@ -196,12 +196,25 @@ static int make_secure_pte(pte_t *ptep, unsigned long addr,
> if (!page_ref_freeze(page, expected))
> return -EBUSY;
> set_bit(PG_arch_1, &page->flags);
> - rc = uv_call(0, (u64)uvcb);
> + /*
> + * If the UVC does not succeed or fail immediately, we don't want to
> + * loop for long, or we might get stall notifications.
> + * On the other hand, this is a complex scenario and we are holding a lot of
> + * locks, so we can't easily sleep and reschedule. We try only once,
> + * and if the UVC returned busy or partial completion, we return
> + * -EAGAIN and we let the callers deal with it.
> + */
> + cc = __uv_call(0, (u64)uvcb);
> page_ref_unfreeze(page, expected);
> - /* Return -ENXIO if the page was not mapped, -EINVAL otherwise */
> - if (rc)
> - rc = uvcb->rc == 0x10a ? -ENXIO : -EINVAL;
> - return rc;
> + /*
> + * Return -ENXIO if the page was not mapped, -EINVAL for other errors.
> + * If busy or partially completed, return -EAGAIN.
> + */
> + if (cc == UVC_CC_OK)
> + return 0;
> + else if (cc == UVC_CC_BUSY || cc == UVC_CC_PARTIAL)
> + return -EAGAIN;
> + return uvcb->rc == 0x10a ? -ENXIO : -EINVAL;
> }
>
> /*
> @@ -254,6 +267,10 @@ int gmap_make_secure(struct gmap *gmap, unsigned long gaddr, void *uvcb)
> mmap_read_unlock(gmap->mm);
>
> if (rc == -EAGAIN) {
> + /*
> + * If we are here because the UVC returned busy or partial
> + * completion, this is just a useless check, but it is safe.
> + */
> wait_on_page_writeback(page);
> } else if (rc == -EBUSY) {
> /*
> diff --git a/arch/s390/kvm/intercept.c b/arch/s390/kvm/intercept.c
> index 72b25b7cc6ae..47833ade4da5 100644
> --- a/arch/s390/kvm/intercept.c
> +++ b/arch/s390/kvm/intercept.c
> @@ -516,6 +516,11 @@ static int handle_pv_uvc(struct kvm_vcpu *vcpu)
> */
> if (rc == -EINVAL)
> return 0;
> + /*
> + * If we got -EAGAIN here, we simply return it. It will eventually
> + * get propagated all the way to userspace, which should then try
> + * again.
> + */
This cpoment is new over v4, right? Can this happen often? If not then this is ok
otherwise we should consider your proposal of doing
if (rc == -EINVAL || rc == -EAGAIN)
to reduce overhead.
Anyway,for both ways
Reviewed-by: Christian Borntraeger <[email protected]>
> return rc;
> }
>
>
On Wed, 6 Oct 2021 17:54:00 +0200
Christian Borntraeger <[email protected]> wrote:
> Am 20.09.21 um 15:24 schrieb Claudio Imbrenda:
> > Improve make_secure_pte to avoid stalls when the system is heavily
> > overcommitted. This was especially problematic in kvm_s390_pv_unpack,
> > because of the loop over all pages that needed unpacking.
> >
> > Due to the locks being held, it was not possible to simply replace
> > uv_call with uv_call_sched. A more complex approach was
> > needed, in which uv_call is replaced with __uv_call, which does not
> > loop. When the UVC needs to be executed again, -EAGAIN is returned, and
> > the caller (or its caller) will try again.
> >
> > When -EAGAIN is returned, the path is the same as when the page is in
> > writeback (and the writeback check is also performed, which is
> > harmless).
> >
> > Signed-off-by: Claudio Imbrenda <[email protected]>
> > Fixes: 214d9bbcd3a672 ("s390/mm: provide memory management functions for protected KVM guests")
> > ---
> > arch/s390/kernel/uv.c | 29 +++++++++++++++++++++++------
> > arch/s390/kvm/intercept.c | 5 +++++
> > 2 files changed, 28 insertions(+), 6 deletions(-)
> >
> > diff --git a/arch/s390/kernel/uv.c b/arch/s390/kernel/uv.c
> > index aeb0a15bcbb7..68a8fbafcb9c 100644
> > --- a/arch/s390/kernel/uv.c
> > +++ b/arch/s390/kernel/uv.c
> > @@ -180,7 +180,7 @@ static int make_secure_pte(pte_t *ptep, unsigned long addr,
> > {
> > pte_t entry = READ_ONCE(*ptep);
> > struct page *page;
> > - int expected, rc = 0;
> > + int expected, cc = 0;
> >
> > if (!pte_present(entry))
> > return -ENXIO;
> > @@ -196,12 +196,25 @@ static int make_secure_pte(pte_t *ptep, unsigned long addr,
> > if (!page_ref_freeze(page, expected))
> > return -EBUSY;
> > set_bit(PG_arch_1, &page->flags);
> > - rc = uv_call(0, (u64)uvcb);
> > + /*
> > + * If the UVC does not succeed or fail immediately, we don't want to
> > + * loop for long, or we might get stall notifications.
> > + * On the other hand, this is a complex scenario and we are holding a lot of
> > + * locks, so we can't easily sleep and reschedule. We try only once,
> > + * and if the UVC returned busy or partial completion, we return
> > + * -EAGAIN and we let the callers deal with it.
> > + */
> > + cc = __uv_call(0, (u64)uvcb);
> > page_ref_unfreeze(page, expected);
> > - /* Return -ENXIO if the page was not mapped, -EINVAL otherwise */
> > - if (rc)
> > - rc = uvcb->rc == 0x10a ? -ENXIO : -EINVAL;
> > - return rc;
> > + /*
> > + * Return -ENXIO if the page was not mapped, -EINVAL for other errors.
> > + * If busy or partially completed, return -EAGAIN.
> > + */
> > + if (cc == UVC_CC_OK)
> > + return 0;
> > + else if (cc == UVC_CC_BUSY || cc == UVC_CC_PARTIAL)
> > + return -EAGAIN;
> > + return uvcb->rc == 0x10a ? -ENXIO : -EINVAL;
> > }
> >
> > /*
> > @@ -254,6 +267,10 @@ int gmap_make_secure(struct gmap *gmap, unsigned long gaddr, void *uvcb)
> > mmap_read_unlock(gmap->mm);
> >
> > if (rc == -EAGAIN) {
> > + /*
> > + * If we are here because the UVC returned busy or partial
> > + * completion, this is just a useless check, but it is safe.
> > + */
> > wait_on_page_writeback(page);
> > } else if (rc == -EBUSY) {
> > /*
> > diff --git a/arch/s390/kvm/intercept.c b/arch/s390/kvm/intercept.c
> > index 72b25b7cc6ae..47833ade4da5 100644
> > --- a/arch/s390/kvm/intercept.c
> > +++ b/arch/s390/kvm/intercept.c
> > @@ -516,6 +516,11 @@ static int handle_pv_uvc(struct kvm_vcpu *vcpu)
> > */
> > if (rc == -EINVAL)
> > return 0;
> > + /*
> > + * If we got -EAGAIN here, we simply return it. It will eventually
> > + * get propagated all the way to userspace, which should then try
> > + * again.
> > + */
>
> This cpoment is new over v4, right? Can this happen often? If not then this is ok
> otherwise we should consider your proposal of doing
yes, the comment is new. I would expect this to happen only when the
system is under load. in any case this is better than busy waiting on
the UVC
>
> if (rc == -EINVAL || rc == -EAGAIN)
>
> to reduce overhead.
>
> Anyway,for both ways
>
> Reviewed-by: Christian Borntraeger <[email protected]>
>
> > return rc;
> > }
> >
> >
On 9/20/21 15:24, Claudio Imbrenda wrote:
> Improve make_secure_pte to avoid stalls when the system is heavily
> overcommitted. This was especially problematic in kvm_s390_pv_unpack,
> because of the loop over all pages that needed unpacking.
>
> Due to the locks being held, it was not possible to simply replace
> uv_call with uv_call_sched. A more complex approach was
> needed, in which uv_call is replaced with __uv_call, which does not
> loop. When the UVC needs to be executed again, -EAGAIN is returned, and
> the caller (or its caller) will try again.
>
> When -EAGAIN is returned, the path is the same as when the page is in
> writeback (and the writeback check is also performed, which is
> harmless).
>
> Signed-off-by: Claudio Imbrenda <[email protected]>
> Fixes: 214d9bbcd3a672 ("s390/mm: provide memory management functions for protected KVM guests")
Reviewed-by: Janosch Frank <[email protected]>
> ---
> arch/s390/kernel/uv.c | 29 +++++++++++++++++++++++------
> arch/s390/kvm/intercept.c | 5 +++++
> 2 files changed, 28 insertions(+), 6 deletions(-)
>
> diff --git a/arch/s390/kernel/uv.c b/arch/s390/kernel/uv.c
> index aeb0a15bcbb7..68a8fbafcb9c 100644
> --- a/arch/s390/kernel/uv.c
> +++ b/arch/s390/kernel/uv.c
> @@ -180,7 +180,7 @@ static int make_secure_pte(pte_t *ptep, unsigned long addr,
> {
> pte_t entry = READ_ONCE(*ptep);
> struct page *page;
> - int expected, rc = 0;
> + int expected, cc = 0;
>
> if (!pte_present(entry))
> return -ENXIO;
> @@ -196,12 +196,25 @@ static int make_secure_pte(pte_t *ptep, unsigned long addr,
> if (!page_ref_freeze(page, expected))
> return -EBUSY;
> set_bit(PG_arch_1, &page->flags);
> - rc = uv_call(0, (u64)uvcb);
> + /*
> + * If the UVC does not succeed or fail immediately, we don't want to
> + * loop for long, or we might get stall notifications.
> + * On the other hand, this is a complex scenario and we are holding a lot of
> + * locks, so we can't easily sleep and reschedule. We try only once,
> + * and if the UVC returned busy or partial completion, we return
> + * -EAGAIN and we let the callers deal with it.
> + */
> + cc = __uv_call(0, (u64)uvcb);
> page_ref_unfreeze(page, expected);
> - /* Return -ENXIO if the page was not mapped, -EINVAL otherwise */
> - if (rc)
> - rc = uvcb->rc == 0x10a ? -ENXIO : -EINVAL;
> - return rc;
> + /*
> + * Return -ENXIO if the page was not mapped, -EINVAL for other errors.
> + * If busy or partially completed, return -EAGAIN.
> + */
> + if (cc == UVC_CC_OK)
> + return 0;
> + else if (cc == UVC_CC_BUSY || cc == UVC_CC_PARTIAL)
> + return -EAGAIN;
> + return uvcb->rc == 0x10a ? -ENXIO : -EINVAL;
> }
>
> /*
> @@ -254,6 +267,10 @@ int gmap_make_secure(struct gmap *gmap, unsigned long gaddr, void *uvcb)
> mmap_read_unlock(gmap->mm);
>
> if (rc == -EAGAIN) {
> + /*
> + * If we are here because the UVC returned busy or partial
> + * completion, this is just a useless check, but it is safe.
> + */
> wait_on_page_writeback(page);
> } else if (rc == -EBUSY) {
> /*
> diff --git a/arch/s390/kvm/intercept.c b/arch/s390/kvm/intercept.c
> index 72b25b7cc6ae..47833ade4da5 100644
> --- a/arch/s390/kvm/intercept.c
> +++ b/arch/s390/kvm/intercept.c
> @@ -516,6 +516,11 @@ static int handle_pv_uvc(struct kvm_vcpu *vcpu)
> */
> if (rc == -EINVAL)
> return 0;
> + /*
> + * If we got -EAGAIN here, we simply return it. It will eventually
> + * get propagated all the way to userspace, which should then try
> + * again.
> + */
> return rc;
> }
>
>
Am 20.09.21 um 15:24 schrieb Claudio Imbrenda:
> Improve make_secure_pte to avoid stalls when the system is heavily
> overcommitted. This was especially problematic in kvm_s390_pv_unpack,
> because of the loop over all pages that needed unpacking.
>
> Due to the locks being held, it was not possible to simply replace
> uv_call with uv_call_sched. A more complex approach was
> needed, in which uv_call is replaced with __uv_call, which does not
> loop. When the UVC needs to be executed again, -EAGAIN is returned, and
> the caller (or its caller) will try again.
>
> When -EAGAIN is returned, the path is the same as when the page is in
> writeback (and the writeback check is also performed, which is
> harmless).
>
> Signed-off-by: Claudio Imbrenda <[email protected]>
> Fixes: 214d9bbcd3a672 ("s390/mm: provide memory management functions for protected KVM guests")
applied this one.
> ---
> arch/s390/kernel/uv.c | 29 +++++++++++++++++++++++------
> arch/s390/kvm/intercept.c | 5 +++++
> 2 files changed, 28 insertions(+), 6 deletions(-)
>
> diff --git a/arch/s390/kernel/uv.c b/arch/s390/kernel/uv.c
> index aeb0a15bcbb7..68a8fbafcb9c 100644
> --- a/arch/s390/kernel/uv.c
> +++ b/arch/s390/kernel/uv.c
> @@ -180,7 +180,7 @@ static int make_secure_pte(pte_t *ptep, unsigned long addr,
> {
> pte_t entry = READ_ONCE(*ptep);
> struct page *page;
> - int expected, rc = 0;
> + int expected, cc = 0;
>
> if (!pte_present(entry))
> return -ENXIO;
> @@ -196,12 +196,25 @@ static int make_secure_pte(pte_t *ptep, unsigned long addr,
> if (!page_ref_freeze(page, expected))
> return -EBUSY;
> set_bit(PG_arch_1, &page->flags);
> - rc = uv_call(0, (u64)uvcb);
> + /*
> + * If the UVC does not succeed or fail immediately, we don't want to
> + * loop for long, or we might get stall notifications.
> + * On the other hand, this is a complex scenario and we are holding a lot of
> + * locks, so we can't easily sleep and reschedule. We try only once,
> + * and if the UVC returned busy or partial completion, we return
> + * -EAGAIN and we let the callers deal with it.
> + */
> + cc = __uv_call(0, (u64)uvcb);
> page_ref_unfreeze(page, expected);
> - /* Return -ENXIO if the page was not mapped, -EINVAL otherwise */
> - if (rc)
> - rc = uvcb->rc == 0x10a ? -ENXIO : -EINVAL;
> - return rc;
> + /*
> + * Return -ENXIO if the page was not mapped, -EINVAL for other errors.
> + * If busy or partially completed, return -EAGAIN.
> + */
> + if (cc == UVC_CC_OK)
> + return 0;
> + else if (cc == UVC_CC_BUSY || cc == UVC_CC_PARTIAL)
> + return -EAGAIN;
> + return uvcb->rc == 0x10a ? -ENXIO : -EINVAL;
> }
>
> /*
> @@ -254,6 +267,10 @@ int gmap_make_secure(struct gmap *gmap, unsigned long gaddr, void *uvcb)
> mmap_read_unlock(gmap->mm);
>
> if (rc == -EAGAIN) {
> + /*
> + * If we are here because the UVC returned busy or partial
> + * completion, this is just a useless check, but it is safe.
> + */
> wait_on_page_writeback(page);
> } else if (rc == -EBUSY) {
> /*
> diff --git a/arch/s390/kvm/intercept.c b/arch/s390/kvm/intercept.c
> index 72b25b7cc6ae..47833ade4da5 100644
> --- a/arch/s390/kvm/intercept.c
> +++ b/arch/s390/kvm/intercept.c
> @@ -516,6 +516,11 @@ static int handle_pv_uvc(struct kvm_vcpu *vcpu)
> */
> if (rc == -EINVAL)
> return 0;
> + /*
> + * If we got -EAGAIN here, we simply return it. It will eventually
> + * get propagated all the way to userspace, which should then try
> + * again.
> + */
> return rc;
> }
>
>