From: Alexander Graf <agraf@suse.de> To: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: kvm@vger.kernel.org, kvm-ppc <kvm-ppc@vger.kernel.org> Subject: Re: [PATCH 2/2] pseries: Correctly create ibm,segment-page-sizes property Date: Thu, 10 May 2012 19:57:23 +0200 [thread overview] Message-ID: <4FAC0183.2020602@suse.de> (raw) In-Reply-To: <1335505904.4578.10.camel@pasglop> On 04/27/2012 07:51 AM, Benjamin Herrenschmidt wrote: > The core tcg/kvm code for ppc64 now has at least the outline > capability to support pagesizes beyond the standard 4k and 16MB. The > CPUState is initialized with information advertising the available > pagesizes and their correct encodings, and under the right KVM setup > this will be populated with page sizes beyond the standard. > > Obviously guests can't use the extra page sizes unless they know > they're present. For the pseries machine, at least, there is a > defined method for conveying exactly this information, the > "ibm-segment-page-sizes" property in the guest device tree. > > This patch generates this property using the supported page size > information that's already in the CPUState. > > Signed-off-by: Nishanth Aravamudan<nacc@us.ibm.com> > Signed-off-by: David Gibson<david@gibson.dropbear.id.au> > Signed-off-by: Benjamin Herrenschmidt<benh@kernel.crashing.org> > --- > hw/spapr.c | 42 ++++++++++++++++++++++++++++++++++++++++++ > target-ppc/kvm.c | 11 +++++++++-- > 2 files changed, 51 insertions(+), 2 deletions(-) > > diff --git a/hw/spapr.c b/hw/spapr.c > index 94a4e1e..7c36903 100644 > --- a/hw/spapr.c > +++ b/hw/spapr.c > @@ -149,6 +149,39 @@ static int spapr_set_associativity(void *fdt, sPAPREnvironment *spapr) > return ret; > } > > + > +static size_t create_page_sizes_prop(CPUPPCState *env, uint32_t *prop, > + size_t maxsize) > +{ > + size_t maxcells = maxsize / sizeof(uint32_t); > + int i, j, count; > + uint32_t *p = prop; > + > + for (i = 0; i< PPC_PAGE_SIZES_MAX_SZ; i++) { > + struct ppc_one_seg_page_size *sps =&env->sps.sps[i]; > + > + if (!sps->page_shift) { > + break; > + } > + for (count = 0; count< PPC_PAGE_SIZES_MAX_SZ; count++) { > + if (sps->enc[count].page_shift == 0) { > + break; > + } > + } > + if ((p - prop)>= (maxcells - 3 - count * 2)) Is this valid C? Can you substract one pointer from another and compare the result with an int? > + break; Braces? Please run checkpatch :) > + *(p++) = cpu_to_be32(sps->page_shift); > + *(p++) = cpu_to_be32(sps->slb_enc); > + *(p++) = cpu_to_be32(count); > + for (j = 0; j< count; j++) { > + *(p++) = cpu_to_be32(sps->enc[j].page_shift); > + *(p++) = cpu_to_be32(sps->enc[j].pte_enc); > + } > + } > + > + return (p - prop) * sizeof(uint32_t); I'd prefer a second integer counter "len" I think :). Pointer arithmentics always make me wary... > +} > + > static void *spapr_create_fdt_skel(const char *cpu_model, > target_phys_addr_t rma_size, > target_phys_addr_t initrd_base, > @@ -304,6 +337,8 @@ static void *spapr_create_fdt_skel(const char *cpu_model, > 0xffffffff, 0xffffffff}; > uint32_t tbfreq = kvm_enabled() ? kvmppc_get_tbfreq() : TIMEBASE_FREQ; > uint32_t cpufreq = kvm_enabled() ? kvmppc_get_clockfreq() : 1000000000; > + uint32_t page_sizes_prop[64]; > + size_t page_sizes_prop_size; > > if ((index % smt) != 0) { > continue; > @@ -368,6 +403,13 @@ static void *spapr_create_fdt_skel(const char *cpu_model, > _FDT((fdt_property_cell(fdt, "ibm,dfp", 1))); > } > > + page_sizes_prop_size = create_page_sizes_prop(env, page_sizes_prop, > + sizeof(page_sizes_prop)); > + if (page_sizes_prop_size) { > + _FDT((fdt_property(fdt, "ibm,segment-page-sizes", > + page_sizes_prop, page_sizes_prop_size))); > + } > + > _FDT((fdt_end_node(fdt))); > } > > diff --git a/target-ppc/kvm.c b/target-ppc/kvm.c > index 77aa186..860711c 100644 > --- a/target-ppc/kvm.c > +++ b/target-ppc/kvm.c > @@ -201,6 +201,7 @@ static void kvm_get_fallback_smmu_info(CPUPPCState *env, > if (kvm_check_extension(env->kvm_state, KVM_CAP_PPC_GET_PVINFO)) { > /* No flags */ > info->flags = 0; > + info->slb_size = 64; Eh - this one belongs in the first patch, no? > > /* Standard 4k base page size segment */ > info->sps[0].page_shift = 12; > @@ -218,9 +219,15 @@ static void kvm_get_fallback_smmu_info(CPUPPCState *env, > > /* HV KVM has backing store size restrictions */ > info->flags = KVM_PPC_PAGE_SIZES_REAL; > + if (env->mmu_model == POWERPC_MMU_2_06) { > + info->slb_size = 32; > + } else { > + info->slb_size = 64; > + } This assumes that we're always using -cpu host. Is there any more reliable way of calculating the slb size? Otherwise maybe we should just error out in non-cpu-host cases for HV mode. > > - if (env->mmu_model& POWERPC_MMU_1TSEG) > - info->flags = KVM_PPC_1T_SEGMENTS; > + if (env->mmu_model& POWERPC_MMU_1TSEG) { > + info->flags |= KVM_PPC_1T_SEGMENTS; > + } Ahem :) Alex
WARNING: multiple messages have this Message-ID (diff)
From: Alexander Graf <agraf@suse.de> To: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: kvm@vger.kernel.org, kvm-ppc <kvm-ppc@vger.kernel.org> Subject: Re: [PATCH 2/2] pseries: Correctly create ibm,segment-page-sizes property Date: Thu, 10 May 2012 17:57:23 +0000 [thread overview] Message-ID: <4FAC0183.2020602@suse.de> (raw) In-Reply-To: <1335505904.4578.10.camel@pasglop> On 04/27/2012 07:51 AM, Benjamin Herrenschmidt wrote: > The core tcg/kvm code for ppc64 now has at least the outline > capability to support pagesizes beyond the standard 4k and 16MB. The > CPUState is initialized with information advertising the available > pagesizes and their correct encodings, and under the right KVM setup > this will be populated with page sizes beyond the standard. > > Obviously guests can't use the extra page sizes unless they know > they're present. For the pseries machine, at least, there is a > defined method for conveying exactly this information, the > "ibm-segment-page-sizes" property in the guest device tree. > > This patch generates this property using the supported page size > information that's already in the CPUState. > > Signed-off-by: Nishanth Aravamudan<nacc@us.ibm.com> > Signed-off-by: David Gibson<david@gibson.dropbear.id.au> > Signed-off-by: Benjamin Herrenschmidt<benh@kernel.crashing.org> > --- > hw/spapr.c | 42 ++++++++++++++++++++++++++++++++++++++++++ > target-ppc/kvm.c | 11 +++++++++-- > 2 files changed, 51 insertions(+), 2 deletions(-) > > diff --git a/hw/spapr.c b/hw/spapr.c > index 94a4e1e..7c36903 100644 > --- a/hw/spapr.c > +++ b/hw/spapr.c > @@ -149,6 +149,39 @@ static int spapr_set_associativity(void *fdt, sPAPREnvironment *spapr) > return ret; > } > > + > +static size_t create_page_sizes_prop(CPUPPCState *env, uint32_t *prop, > + size_t maxsize) > +{ > + size_t maxcells = maxsize / sizeof(uint32_t); > + int i, j, count; > + uint32_t *p = prop; > + > + for (i = 0; i< PPC_PAGE_SIZES_MAX_SZ; i++) { > + struct ppc_one_seg_page_size *sps =&env->sps.sps[i]; > + > + if (!sps->page_shift) { > + break; > + } > + for (count = 0; count< PPC_PAGE_SIZES_MAX_SZ; count++) { > + if (sps->enc[count].page_shift = 0) { > + break; > + } > + } > + if ((p - prop)>= (maxcells - 3 - count * 2)) Is this valid C? Can you substract one pointer from another and compare the result with an int? > + break; Braces? Please run checkpatch :) > + *(p++) = cpu_to_be32(sps->page_shift); > + *(p++) = cpu_to_be32(sps->slb_enc); > + *(p++) = cpu_to_be32(count); > + for (j = 0; j< count; j++) { > + *(p++) = cpu_to_be32(sps->enc[j].page_shift); > + *(p++) = cpu_to_be32(sps->enc[j].pte_enc); > + } > + } > + > + return (p - prop) * sizeof(uint32_t); I'd prefer a second integer counter "len" I think :). Pointer arithmentics always make me wary... > +} > + > static void *spapr_create_fdt_skel(const char *cpu_model, > target_phys_addr_t rma_size, > target_phys_addr_t initrd_base, > @@ -304,6 +337,8 @@ static void *spapr_create_fdt_skel(const char *cpu_model, > 0xffffffff, 0xffffffff}; > uint32_t tbfreq = kvm_enabled() ? kvmppc_get_tbfreq() : TIMEBASE_FREQ; > uint32_t cpufreq = kvm_enabled() ? kvmppc_get_clockfreq() : 1000000000; > + uint32_t page_sizes_prop[64]; > + size_t page_sizes_prop_size; > > if ((index % smt) != 0) { > continue; > @@ -368,6 +403,13 @@ static void *spapr_create_fdt_skel(const char *cpu_model, > _FDT((fdt_property_cell(fdt, "ibm,dfp", 1))); > } > > + page_sizes_prop_size = create_page_sizes_prop(env, page_sizes_prop, > + sizeof(page_sizes_prop)); > + if (page_sizes_prop_size) { > + _FDT((fdt_property(fdt, "ibm,segment-page-sizes", > + page_sizes_prop, page_sizes_prop_size))); > + } > + > _FDT((fdt_end_node(fdt))); > } > > diff --git a/target-ppc/kvm.c b/target-ppc/kvm.c > index 77aa186..860711c 100644 > --- a/target-ppc/kvm.c > +++ b/target-ppc/kvm.c > @@ -201,6 +201,7 @@ static void kvm_get_fallback_smmu_info(CPUPPCState *env, > if (kvm_check_extension(env->kvm_state, KVM_CAP_PPC_GET_PVINFO)) { > /* No flags */ > info->flags = 0; > + info->slb_size = 64; Eh - this one belongs in the first patch, no? > > /* Standard 4k base page size segment */ > info->sps[0].page_shift = 12; > @@ -218,9 +219,15 @@ static void kvm_get_fallback_smmu_info(CPUPPCState *env, > > /* HV KVM has backing store size restrictions */ > info->flags = KVM_PPC_PAGE_SIZES_REAL; > + if (env->mmu_model = POWERPC_MMU_2_06) { > + info->slb_size = 32; > + } else { > + info->slb_size = 64; > + } This assumes that we're always using -cpu host. Is there any more reliable way of calculating the slb size? Otherwise maybe we should just error out in non-cpu-host cases for HV mode. > > - if (env->mmu_model& POWERPC_MMU_1TSEG) > - info->flags = KVM_PPC_1T_SEGMENTS; > + if (env->mmu_model& POWERPC_MMU_1TSEG) { > + info->flags |= KVM_PPC_1T_SEGMENTS; > + } Ahem :) Alex
next prev parent reply other threads:[~2012-05-10 17:57 UTC|newest] Thread overview: 27+ messages / expand[flat|nested] mbox.gz Atom feed top 2012-04-27 5:43 [PATCH] kvm/powerpc: Add new ioctl to retreive server MMU infos Benjamin Herrenschmidt 2012-04-27 5:43 ` Benjamin Herrenschmidt 2012-04-27 5:51 ` [PATCH 1/2] ppc64: Rudimentary Support for extra page sizes on server CPUs Benjamin Herrenschmidt 2012-04-27 5:51 ` Benjamin Herrenschmidt 2012-05-10 17:49 ` Alexander Graf 2012-05-10 17:49 ` Alexander Graf 2012-05-10 17:55 ` Avi Kivity 2012-05-10 17:55 ` Avi Kivity 2012-05-10 21:43 ` Benjamin Herrenschmidt 2012-05-10 21:43 ` Benjamin Herrenschmidt 2012-04-27 5:51 ` [PATCH 2/2] pseries: Correctly create ibm,segment-page-sizes property Benjamin Herrenschmidt 2012-04-27 5:51 ` Benjamin Herrenschmidt 2012-05-10 17:57 ` Alexander Graf [this message] 2012-05-10 17:57 ` Alexander Graf 2012-05-10 21:49 ` Benjamin Herrenschmidt 2012-05-10 21:49 ` Benjamin Herrenschmidt 2012-05-03 8:09 ` [PATCH] kvm/powerpc: Add new ioctl to retreive server MMU infos Alexander Graf 2012-05-03 8:09 ` Alexander Graf 2012-06-19 5:56 [PATCH 1/2] ppc64: Rudimentary Support for extra page sizes on server CPUs Benjamin Herrenschmidt 2012-06-19 5:56 ` Benjamin Herrenschmidt 2012-06-19 5:56 ` [Qemu-devel] " Benjamin Herrenschmidt 2012-06-19 20:57 ` Alexander Graf 2012-06-19 20:57 ` Alexander Graf 2012-06-19 20:57 ` [Qemu-devel] " Alexander Graf 2012-06-19 5:56 [PATCH 2/2] pseries: Correctly create ibm,segment-page-sizes property Benjamin Herrenschmidt 2012-06-19 5:56 ` Benjamin Herrenschmidt 2012-06-19 5:56 ` [Qemu-devel] [PATCH 2/2] pseries: Correctly create ibm, segment-page-sizes property Benjamin Herrenschmidt
Reply instructions: You may reply publicly to this message via plain-text email using any one of the following methods: * Save the following mbox file, import it into your mail client, and reply-to-all from there: mbox Avoid top-posting and favor interleaved quoting: https://en.wikipedia.org/wiki/Posting_style#Interleaved_style * Reply using the --to, --cc, and --in-reply-to switches of git-send-email(1): git send-email \ --in-reply-to=4FAC0183.2020602@suse.de \ --to=agraf@suse.de \ --cc=benh@kernel.crashing.org \ --cc=kvm-ppc@vger.kernel.org \ --cc=kvm@vger.kernel.org \ /path/to/YOUR_REPLY https://kernel.org/pub/software/scm/git/docs/git-send-email.html * If your mail client supports setting the In-Reply-To header via mailto: links, try the mailto: linkBe sure your reply has a Subject: header at the top and a blank line before the message body.
This is an external index of several public inboxes, see mirroring instructions on how to clone and mirror all data and code used by this external index.