| Message ID | 20210920132502.36111-5-imbrenda@linux.ibm.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | KVM: s390: pv: implement lazy destroy for reboot | expand |
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 <imbrenda@linux.ibm.com> > 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 <borntraeger@de.ibm.com> > return rc; > } > >
On Wed, 6 Oct 2021 17:54:00 +0200 Christian Borntraeger <borntraeger@de.ibm.com> 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 <imbrenda@linux.ibm.com> > > 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 <borntraeger@de.ibm.com> > > > 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 <imbrenda@linux.ibm.com> > Fixes: 214d9bbcd3a672 ("s390/mm: provide memory management functions for protected KVM guests") Reviewed-by: Janosch Frank <frankja@linux.ibm.com> > --- > 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 <imbrenda@linux.ibm.com> > 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; > } > >
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; }
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 <imbrenda@linux.ibm.com> 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(-)