| Message ID | 20210527125134.2116404-4-qperret@google.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | KVM: arm64: Reduce hyp_vmemmap overhead | expand |
On Thu, 27 May 2021 13:51:30 +0100, Quentin Perret <qperret@google.com> wrote: > > The list_head member of struct hyp_page is only needed when the page is > attached to a free-list, which by definition implies the page is free. > As such, nothing prevents us from using the page itself to store the > list_head, hence reducing the size of the vmemmap. > > Signed-off-by: Quentin Perret <qperret@google.com> > --- > arch/arm64/kvm/hyp/include/nvhe/memory.h | 1 - > arch/arm64/kvm/hyp/nvhe/page_alloc.c | 39 ++++++++++++++++++++---- > 2 files changed, 33 insertions(+), 7 deletions(-) > > diff --git a/arch/arm64/kvm/hyp/include/nvhe/memory.h b/arch/arm64/kvm/hyp/include/nvhe/memory.h > index fd78bde939ee..7691ab495eb4 100644 > --- a/arch/arm64/kvm/hyp/include/nvhe/memory.h > +++ b/arch/arm64/kvm/hyp/include/nvhe/memory.h > @@ -12,7 +12,6 @@ struct hyp_page { > unsigned int refcount; > unsigned int order; > struct hyp_pool *pool; > - struct list_head node; > }; > > extern u64 __hyp_vmemmap; > diff --git a/arch/arm64/kvm/hyp/nvhe/page_alloc.c b/arch/arm64/kvm/hyp/nvhe/page_alloc.c > index 7ee882f36767..ce7379f1480b 100644 > --- a/arch/arm64/kvm/hyp/nvhe/page_alloc.c > +++ b/arch/arm64/kvm/hyp/nvhe/page_alloc.c > @@ -62,6 +62,34 @@ static struct hyp_page *__find_buddy_avail(struct hyp_pool *pool, > > } > > +/* > + * Pages that are available for allocation are tracked in free-lists, so we use > + * the pages themselves to store the list nodes to avoid wasting space. As the > + * allocator always returns zeroed pages (which are zeroed on the hyp_put_page() > + * path to optimize allocation speed), we also need to clean-up the list node in > + * each page when we take it out of the list. > + */ > +static inline void page_remove_from_list(struct hyp_page *p) > +{ > + struct list_head *node = (struct list_head *)hyp_page_to_virt(p); Nit: How about changing hyp_page_to_virt() so that it returns a convenient 'void *', and get rid of the ugly casts? > + > + __list_del_entry(node); > + memset(node, 0, sizeof(*node)); > +} > + > +static inline void page_add_to_list(struct hyp_page *p, struct list_head *head) > +{ > + struct list_head *node = (struct list_head *)hyp_page_to_virt(p); > + > + INIT_LIST_HEAD(node); > + list_add_tail(node, head); > +} > + > +static inline struct hyp_page *node_to_page(struct list_head *node) > +{ > + return (struct hyp_page *)hyp_virt_to_page(node); Why is this cast necessary? If I'm not mistaken, hyp_vmemmap is already cast as a 'struct hyp_page *', so hyp_virt_to_page() should return the same type. > +} > + > static void __hyp_attach_page(struct hyp_pool *pool, > struct hyp_page *p) > { > @@ -83,14 +111,14 @@ static void __hyp_attach_page(struct hyp_pool *pool, > break; > > /* Take the buddy out of its list, and coallesce with @p */ > - list_del_init(&buddy->node); > + page_remove_from_list(buddy); > buddy->order = HYP_NO_ORDER; > p = min(p, buddy); > } > > /* Mark the new head, and insert it */ > p->order = order; > - list_add_tail(&p->node, &pool->free_area[order]); > + page_add_to_list(p, &pool->free_area[order]); > } > > static struct hyp_page *__hyp_extract_page(struct hyp_pool *pool, > @@ -99,7 +127,7 @@ static struct hyp_page *__hyp_extract_page(struct hyp_pool *pool, > { > struct hyp_page *buddy; > > - list_del_init(&p->node); > + page_remove_from_list(p); > while (p->order > order) { > /* > * The buddy of order n - 1 currently has HYP_NO_ORDER as it > @@ -110,7 +138,7 @@ static struct hyp_page *__hyp_extract_page(struct hyp_pool *pool, > p->order--; > buddy = __find_buddy_nocheck(pool, p, p->order); > buddy->order = p->order; > - list_add_tail(&buddy->node, &pool->free_area[buddy->order]); > + page_add_to_list(buddy, &pool->free_area[buddy->order]); > } > > return p; > @@ -158,7 +186,7 @@ void *hyp_alloc_pages(struct hyp_pool *pool, unsigned int order) > } > > /* Extract it from the tree at the right order */ > - p = list_first_entry(&pool->free_area[i], struct hyp_page, node); > + p = node_to_page(pool->free_area[i].next); > p = __hyp_extract_page(pool, p, order); > > hyp_set_page_refcounted(p); > @@ -186,7 +214,6 @@ int hyp_pool_init(struct hyp_pool *pool, u64 pfn, unsigned int nr_pages, > for (i = 0; i < nr_pages; i++) { > p[i].pool = pool; > p[i].order = 0; > - INIT_LIST_HEAD(&p[i].node); > hyp_set_page_refcounted(&p[i]); > } > > -- > 2.31.1.818.g46aad6cb9e-goog > > Thanks, M.
On Tuesday 01 Jun 2021 at 15:38:22 (+0100), Marc Zyngier wrote: > On Thu, 27 May 2021 13:51:30 +0100, > Quentin Perret <qperret@google.com> wrote: > > +/* > > + * Pages that are available for allocation are tracked in free-lists, so we use > > + * the pages themselves to store the list nodes to avoid wasting space. As the > > + * allocator always returns zeroed pages (which are zeroed on the hyp_put_page() > > + * path to optimize allocation speed), we also need to clean-up the list node in > > + * each page when we take it out of the list. > > + */ > > +static inline void page_remove_from_list(struct hyp_page *p) > > +{ > > + struct list_head *node = (struct list_head *)hyp_page_to_virt(p); > > Nit: How about changing hyp_page_to_virt() so that it returns a > convenient 'void *', and get rid of the ugly casts? It should already return void *, but I kind of liked the explicit cast here for documentation purpose. We're turning a 'random' piece of unused memory into a typed object, so that felt like a useful annotation. Happy to get rid of it though. > > + > > + __list_del_entry(node); > > + memset(node, 0, sizeof(*node)); > > +} > > + > > +static inline void page_add_to_list(struct hyp_page *p, struct list_head *head) > > +{ > > + struct list_head *node = (struct list_head *)hyp_page_to_virt(p); > > + > > + INIT_LIST_HEAD(node); > > + list_add_tail(node, head); > > +} > > + > > +static inline struct hyp_page *node_to_page(struct list_head *node) > > +{ > > + return (struct hyp_page *)hyp_virt_to_page(node); > > Why is this cast necessary? If I'm not mistaken, hyp_vmemmap is > already cast as a 'struct hyp_page *', so hyp_virt_to_page() should > return the same type. Right, that one is totally unnecessary, I'll remove. Cheers, Quentin
On Tue, 01 Jun 2021 16:48:06 +0100, Quentin Perret <qperret@google.com> wrote: > > On Tuesday 01 Jun 2021 at 15:38:22 (+0100), Marc Zyngier wrote: > > On Thu, 27 May 2021 13:51:30 +0100, > > Quentin Perret <qperret@google.com> wrote: > > > +/* > > > + * Pages that are available for allocation are tracked in free-lists, so we use > > > + * the pages themselves to store the list nodes to avoid wasting space. As the > > > + * allocator always returns zeroed pages (which are zeroed on the hyp_put_page() > > > + * path to optimize allocation speed), we also need to clean-up the list node in > > > + * each page when we take it out of the list. > > > + */ > > > +static inline void page_remove_from_list(struct hyp_page *p) > > > +{ > > > + struct list_head *node = (struct list_head *)hyp_page_to_virt(p); > > > > Nit: How about changing hyp_page_to_virt() so that it returns a > > convenient 'void *', and get rid of the ugly casts? > > It should already return void *, but I kind of liked the explicit cast > here for documentation purpose. We're turning a 'random' piece of unused > memory into a typed object, so that felt like a useful annotation. Happy > to get rid of it though. My expectations were that using hyp_page_to_virt() already serves as a pretty big warning that we're doing something unusual. I guess that if we want to be really careful about those, we should then be consistent and make it return a uintptr_t (or unsigned long) instead, actively making use of the cast, consistently, everywhere. Thanks, M.
On Tuesday 01 Jun 2021 at 18:41:49 (+0100), Marc Zyngier wrote: > On Tue, 01 Jun 2021 16:48:06 +0100, > Quentin Perret <qperret@google.com> wrote: > > > > On Tuesday 01 Jun 2021 at 15:38:22 (+0100), Marc Zyngier wrote: > > > On Thu, 27 May 2021 13:51:30 +0100, > > > Quentin Perret <qperret@google.com> wrote: > > > > +/* > > > > + * Pages that are available for allocation are tracked in free-lists, so we use > > > > + * the pages themselves to store the list nodes to avoid wasting space. As the > > > > + * allocator always returns zeroed pages (which are zeroed on the hyp_put_page() > > > > + * path to optimize allocation speed), we also need to clean-up the list node in > > > > + * each page when we take it out of the list. > > > > + */ > > > > +static inline void page_remove_from_list(struct hyp_page *p) > > > > +{ > > > > + struct list_head *node = (struct list_head *)hyp_page_to_virt(p); > > > > > > Nit: How about changing hyp_page_to_virt() so that it returns a > > > convenient 'void *', and get rid of the ugly casts? > > > > It should already return void *, but I kind of liked the explicit cast > > here for documentation purpose. We're turning a 'random' piece of unused > > memory into a typed object, so that felt like a useful annotation. Happy > > to get rid of it though. > > My expectations were that using hyp_page_to_virt() already serves as a > pretty big warning that we're doing something unusual. > > I guess that if we want to be really careful about those, we should > then be consistent and make it return a uintptr_t (or unsigned long) > instead, actively making use of the cast, consistently, everywhere. Right, so I'd prefer keeping it void * for consistency with the EL1 equivalent -- much of this aims to be API compatible with the Linux stuff. I'll get rid of the cast and post a v2 shortly. Cheers, Quentin
diff --git a/arch/arm64/kvm/hyp/include/nvhe/memory.h b/arch/arm64/kvm/hyp/include/nvhe/memory.h index fd78bde939ee..7691ab495eb4 100644 --- a/arch/arm64/kvm/hyp/include/nvhe/memory.h +++ b/arch/arm64/kvm/hyp/include/nvhe/memory.h @@ -12,7 +12,6 @@ struct hyp_page { unsigned int refcount; unsigned int order; struct hyp_pool *pool; - struct list_head node; }; extern u64 __hyp_vmemmap; diff --git a/arch/arm64/kvm/hyp/nvhe/page_alloc.c b/arch/arm64/kvm/hyp/nvhe/page_alloc.c index 7ee882f36767..ce7379f1480b 100644 --- a/arch/arm64/kvm/hyp/nvhe/page_alloc.c +++ b/arch/arm64/kvm/hyp/nvhe/page_alloc.c @@ -62,6 +62,34 @@ static struct hyp_page *__find_buddy_avail(struct hyp_pool *pool, } +/* + * Pages that are available for allocation are tracked in free-lists, so we use + * the pages themselves to store the list nodes to avoid wasting space. As the + * allocator always returns zeroed pages (which are zeroed on the hyp_put_page() + * path to optimize allocation speed), we also need to clean-up the list node in + * each page when we take it out of the list. + */ +static inline void page_remove_from_list(struct hyp_page *p) +{ + struct list_head *node = (struct list_head *)hyp_page_to_virt(p); + + __list_del_entry(node); + memset(node, 0, sizeof(*node)); +} + +static inline void page_add_to_list(struct hyp_page *p, struct list_head *head) +{ + struct list_head *node = (struct list_head *)hyp_page_to_virt(p); + + INIT_LIST_HEAD(node); + list_add_tail(node, head); +} + +static inline struct hyp_page *node_to_page(struct list_head *node) +{ + return (struct hyp_page *)hyp_virt_to_page(node); +} + static void __hyp_attach_page(struct hyp_pool *pool, struct hyp_page *p) { @@ -83,14 +111,14 @@ static void __hyp_attach_page(struct hyp_pool *pool, break; /* Take the buddy out of its list, and coallesce with @p */ - list_del_init(&buddy->node); + page_remove_from_list(buddy); buddy->order = HYP_NO_ORDER; p = min(p, buddy); } /* Mark the new head, and insert it */ p->order = order; - list_add_tail(&p->node, &pool->free_area[order]); + page_add_to_list(p, &pool->free_area[order]); } static struct hyp_page *__hyp_extract_page(struct hyp_pool *pool, @@ -99,7 +127,7 @@ static struct hyp_page *__hyp_extract_page(struct hyp_pool *pool, { struct hyp_page *buddy; - list_del_init(&p->node); + page_remove_from_list(p); while (p->order > order) { /* * The buddy of order n - 1 currently has HYP_NO_ORDER as it @@ -110,7 +138,7 @@ static struct hyp_page *__hyp_extract_page(struct hyp_pool *pool, p->order--; buddy = __find_buddy_nocheck(pool, p, p->order); buddy->order = p->order; - list_add_tail(&buddy->node, &pool->free_area[buddy->order]); + page_add_to_list(buddy, &pool->free_area[buddy->order]); } return p; @@ -158,7 +186,7 @@ void *hyp_alloc_pages(struct hyp_pool *pool, unsigned int order) } /* Extract it from the tree at the right order */ - p = list_first_entry(&pool->free_area[i], struct hyp_page, node); + p = node_to_page(pool->free_area[i].next); p = __hyp_extract_page(pool, p, order); hyp_set_page_refcounted(p); @@ -186,7 +214,6 @@ int hyp_pool_init(struct hyp_pool *pool, u64 pfn, unsigned int nr_pages, for (i = 0; i < nr_pages; i++) { p[i].pool = pool; p[i].order = 0; - INIT_LIST_HEAD(&p[i].node); hyp_set_page_refcounted(&p[i]); }
The list_head member of struct hyp_page is only needed when the page is attached to a free-list, which by definition implies the page is free. As such, nothing prevents us from using the page itself to store the list_head, hence reducing the size of the vmemmap. Signed-off-by: Quentin Perret <qperret@google.com> --- arch/arm64/kvm/hyp/include/nvhe/memory.h | 1 - arch/arm64/kvm/hyp/nvhe/page_alloc.c | 39 ++++++++++++++++++++---- 2 files changed, 33 insertions(+), 7 deletions(-)