Message ID | 20190514122456.28559-15-julien.grall@arm.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | xen/arm: Clean-up & fixes in boot/mm code | expand |
On Tue, 14 May 2019, Julien Grall wrote: > The page-table walker is configured to use the same shareability and > cacheability as the access performed when updating the page-tables. This > means cleaning the cache for secondary CPUs runtime page-tables is > unnecessary. All right. Is there an explicit configuration for the shareability and cacheability used by the page-table walker or is it specified as such in the Arm Arm? Also, isn't it possible that CPUs on a different cluster (big.LITTLE) would have issues with this if the cache could be split between the two clusters? > Signed-off-by: Julien Grall <julien.grall@arm.com> > Reviewed-by: Andrii Anisov <andrii_anisov@epam.com> > > --- > Changes in v2: > - Add Andrii's reviewed-by > --- > xen/arch/arm/mm.c | 3 --- > 1 file changed, 3 deletions(-) > > diff --git a/xen/arch/arm/mm.c b/xen/arch/arm/mm.c > index cda2847d00..6db7dda0da 100644 > --- a/xen/arch/arm/mm.c > +++ b/xen/arch/arm/mm.c > @@ -769,9 +769,6 @@ int init_secondary_pagetables(int cpu) > write_pte(&first[first_table_offset(DOMHEAP_VIRT_START+i*FIRST_SIZE)], pte); > } > > - clean_dcache_va_range(first, PAGE_SIZE); > - clean_dcache_va_range(domheap, DOMHEAP_SECOND_PAGES*PAGE_SIZE); > - > per_cpu(xen_pgtable, cpu) = first; > per_cpu(xen_dommap, cpu) = domheap; > > -- > 2.11.0 >
On 05/06/2019 00:11, Stefano Stabellini wrote: > On Tue, 14 May 2019, Julien Grall wrote: >> The page-table walker is configured to use the same shareability and >> cacheability as the access performed when updating the page-tables. This >> means cleaning the cache for secondary CPUs runtime page-tables is >> unnecessary. > > All right. Is there an explicit configuration for the shareability and > cacheability used by the page-table walker or is it specified as such in > the Arm Arm? See the configuration of TCR_EL2, I can mention it. > Also, isn't it possible that CPUs on a different cluster > (big.LITTLE) would have issues with this if the cache could be split > between the two clusters? I don't understand this... Cache should be coherent when a CPU leaves EL3. But we already share some bits of the page tables between the processor (see create_xen_page_tables). So I don't see where there is a possible problem here. Cheers, > > >> Signed-off-by: Julien Grall <julien.grall@arm.com> >> Reviewed-by: Andrii Anisov <andrii_anisov@epam.com> >> >> --- >> Changes in v2: >> - Add Andrii's reviewed-by >> --- >> xen/arch/arm/mm.c | 3 --- >> 1 file changed, 3 deletions(-) >> >> diff --git a/xen/arch/arm/mm.c b/xen/arch/arm/mm.c >> index cda2847d00..6db7dda0da 100644 >> --- a/xen/arch/arm/mm.c >> +++ b/xen/arch/arm/mm.c >> @@ -769,9 +769,6 @@ int init_secondary_pagetables(int cpu) >> write_pte(&first[first_table_offset(DOMHEAP_VIRT_START+i*FIRST_SIZE)], pte); >> } >> >> - clean_dcache_va_range(first, PAGE_SIZE); >> - clean_dcache_va_range(domheap, DOMHEAP_SECOND_PAGES*PAGE_SIZE); >> - >> per_cpu(xen_pgtable, cpu) = first; >> per_cpu(xen_dommap, cpu) = domheap; >> >> -- >> 2.11.0 >>
Hi, On 05/06/2019 11:19, Julien Grall wrote: > On 05/06/2019 00:11, Stefano Stabellini wrote: >> On Tue, 14 May 2019, Julien Grall wrote: >>> The page-table walker is configured to use the same shareability and >>> cacheability as the access performed when updating the page-tables. This >>> means cleaning the cache for secondary CPUs runtime page-tables is >>> unnecessary. >> >> All right. Is there an explicit configuration for the shareability and >> cacheability used by the page-table walker or is it specified as such in >> the Arm Arm? > > See the configuration of TCR_EL2, I can mention it. > >> Also, isn't it possible that CPUs on a different cluster >> (big.LITTLE) would have issues with this if the cache could be split >> between the two clusters? Gentle ping, can you please clarify your question. > > I don't understand this... Cache should be coherent when a CPU leaves EL3. > But we already share some bits of the page tables between the processor (see > create_xen_page_tables). So I don't see where there is a possible problem here. Cheers,
On Wed, 5 Jun 2019, Julien Grall wrote: > On 05/06/2019 00:11, Stefano Stabellini wrote: > > On Tue, 14 May 2019, Julien Grall wrote: > > > The page-table walker is configured to use the same shareability and > > > cacheability as the access performed when updating the page-tables. This > > > means cleaning the cache for secondary CPUs runtime page-tables is > > > unnecessary. > > > > All right. Is there an explicit configuration for the shareability and > > cacheability used by the page-table walker or is it specified as such in > > the Arm Arm? > > See the configuration of TCR_EL2, I can mention it. That would be nice. I double-checked and it is as you wrote. > > Also, isn't it possible that CPUs on a different cluster > > (big.LITTLE) would have issues with this if the cache could be split > > between the two clusters? > > I don't understand this... Cache should be coherent when a CPU leaves EL3. > But we already share some bits of the page tables between the processor (see > create_xen_page_tables). So I don't see where there is a possible problem > here. If the cache is always coherent across the clusters and the pagetable-walkers of different clusters, then this is fine. > > > Signed-off-by: Julien Grall <julien.grall@arm.com> > > > Reviewed-by: Andrii Anisov <andrii_anisov@epam.com> > > > > > > --- > > > Changes in v2: > > > - Add Andrii's reviewed-by > > > --- > > > xen/arch/arm/mm.c | 3 --- > > > 1 file changed, 3 deletions(-) > > > > > > diff --git a/xen/arch/arm/mm.c b/xen/arch/arm/mm.c > > > index cda2847d00..6db7dda0da 100644 > > > --- a/xen/arch/arm/mm.c > > > +++ b/xen/arch/arm/mm.c > > > @@ -769,9 +769,6 @@ int init_secondary_pagetables(int cpu) > > > write_pte(&first[first_table_offset(DOMHEAP_VIRT_START+i*FIRST_SIZE)], > > > pte); > > > } > > > - clean_dcache_va_range(first, PAGE_SIZE); > > > - clean_dcache_va_range(domheap, DOMHEAP_SECOND_PAGES*PAGE_SIZE); > > > - > > > per_cpu(xen_pgtable, cpu) = first; > > > per_cpu(xen_dommap, cpu) = domheap; > > > > > > -- > > > 2.11.0 > > > > > -- > Julien Grall >
Hi Steano, On 6/10/19 9:28 PM, Stefano Stabellini wrote: > On Wed, 5 Jun 2019, Julien Grall wrote: >> On 05/06/2019 00:11, Stefano Stabellini wrote: >>> On Tue, 14 May 2019, Julien Grall wrote: >>>> The page-table walker is configured to use the same shareability and >>>> cacheability as the access performed when updating the page-tables. This >>>> means cleaning the cache for secondary CPUs runtime page-tables is >>>> unnecessary. >>> >>> All right. Is there an explicit configuration for the shareability and >>> cacheability used by the page-table walker or is it specified as such in >>> the Arm Arm? >> >> See the configuration of TCR_EL2, I can mention it. > > That would be nice. I double-checked and it is as you wrote. Sure. How about: "The page-table walker is configured by TCR_EL2 to use shareability and cacheability as the access performed when updating the page-tables. [...]" > > >>> Also, isn't it possible that CPUs on a different cluster >>> (big.LITTLE) would have issues with this if the cache could be split >>> between the two clusters? >> >> I don't understand this... Cache should be coherent when a CPU leaves EL3. >> But we already share some bits of the page tables between the processor (see >> create_xen_page_tables). So I don't see where there is a possible problem >> here. > > If the cache is always coherent across the clusters and the > pagetable-walkers of different clusters, then this is fine. Xen (and Linux) built on the assumption that all the CPUs (and page-table walker) are in the same shareable domain (i.e innershareable). If you have a platform where it is not the case, then Xen is going to be badly broken. This is also inline with the expectatio from the Arm Arm (B2-123 in DDI 0487D.a): "The Inner Shareable domain is expected to be the set of PEs controlled by a single hypervisor or operating system." Do you have a case where Xen needs to run on PEs in different domains? Cheers,
On Mon, 10 Jun 2019, Julien Grall wrote: > Hi Steano, > > On 6/10/19 9:28 PM, Stefano Stabellini wrote: > > On Wed, 5 Jun 2019, Julien Grall wrote: > > > On 05/06/2019 00:11, Stefano Stabellini wrote: > > > > On Tue, 14 May 2019, Julien Grall wrote: > > > > > The page-table walker is configured to use the same shareability and > > > > > cacheability as the access performed when updating the page-tables. > > > > > This > > > > > means cleaning the cache for secondary CPUs runtime page-tables is > > > > > unnecessary. > > > > > > > > All right. Is there an explicit configuration for the shareability and > > > > cacheability used by the page-table walker or is it specified as such in > > > > the Arm Arm? > > > > > > See the configuration of TCR_EL2, I can mention it. > > That would be nice. I double-checked and it is as you wrote. > > Sure. How about: > > "The page-table walker is configured by TCR_EL2 to use shareability and > cacheability as the access performed when updating the page-tables. [...]" That's great thank you. With that, add my reviewed-by. > > > > Also, isn't it possible that CPUs on a different cluster > > > > (big.LITTLE) would have issues with this if the cache could be split > > > > between the two clusters? > > > > > > I don't understand this... Cache should be coherent when a CPU leaves EL3. > > > But we already share some bits of the page tables between the processor > > > (see > > > create_xen_page_tables). So I don't see where there is a possible problem > > > here. > > > > If the cache is always coherent across the clusters and the > > pagetable-walkers of different clusters, then this is fine. > > Xen (and Linux) built on the assumption that all the CPUs (and page-table > walker) are in the same shareable domain (i.e innershareable). If you have a > platform where it is not the case, then Xen is going to be badly broken. > > This is also inline with the expectatio from the Arm Arm (B2-123 in DDI > 0487D.a): > > "The Inner Shareable domain is expected to be the set of PEs controlled by a > single hypervisor or operating system." > > Do you have a case where Xen needs to run on PEs in different domains? No, thankfully I don't :-) I was worried that some big.LITTLE SoCs might be built like that so. (I don't have any big.LITTLE machines here to confirm/deny.) It is good that we don't have to worry about it.
diff --git a/xen/arch/arm/mm.c b/xen/arch/arm/mm.c index cda2847d00..6db7dda0da 100644 --- a/xen/arch/arm/mm.c +++ b/xen/arch/arm/mm.c @@ -769,9 +769,6 @@ int init_secondary_pagetables(int cpu) write_pte(&first[first_table_offset(DOMHEAP_VIRT_START+i*FIRST_SIZE)], pte); } - clean_dcache_va_range(first, PAGE_SIZE); - clean_dcache_va_range(domheap, DOMHEAP_SECOND_PAGES*PAGE_SIZE); - per_cpu(xen_pgtable, cpu) = first; per_cpu(xen_dommap, cpu) = domheap;