| Message ID | 20141210222912.GA30869@us.ibm.com (mailing list archive) |
|---|---|
| State | Accepted |
| Commit | f34b6c72c3ebaa286d3311a825ef79eccbcca82f |
| Delegated to: | Michael Ellerman |
| Headers | show |
On Wed, 2014-10-12 at 22:29:13 UTC, sukadev@linux.vnet.ibm.com wrote: > Michael Ellerman [mpe@ellerman.id.au] wrote: > | On Tue, 2014-12-09 at 23:06 -0800, Sukadev Bhattiprolu wrote: > | > From 470c16c8955672103a9529c78dffbb239e9e27b8 Mon Sep 17 00:00:00 2001 > | > From: Sukadev Bhattiprolu <sukadev@linux.vnet.ibm.com> > | > Date: Tue, 9 Dec 2014 22:17:46 -0500 > | > Subject: [PATCH 1/2] perf/poweprc/hv-24x7: Use per-cpu page buffer > | > > | > diff --git a/arch/powerpc/perf/hv-24x7.c b/arch/powerpc/perf/hv-24x7.c > | > index dba3408..18e1f49 100644 > | > --- a/arch/powerpc/perf/hv-24x7.c > | > +++ b/arch/powerpc/perf/hv-24x7.c > | > @@ -217,11 +217,14 @@ static bool is_physical_domain(int domain) > | > domain == HV_24X7_PERF_DOMAIN_PHYSICAL_CORE; > | > } > | > > | > +DEFINE_PER_CPU(char, hv_24x7_reqb[4096]); > | > +DEFINE_PER_CPU(char, hv_24x7_resb[4096]); > | > | Do we need it to be 4K aligned also? I would guess so. > > Yes, fixed in the patch below. OK. > | > | Rather than declaring these as char arrays and then casting below, can you pull > | the struct definitions up and then declare the per cpu variables with the > | proper type. > > Well, the structures, used for communication with HV, have variable length > arrays, like: > > struct hv_24x7_request_buffer { > ... > struct hv_24x7_request requests[]; > }; > > i.e the buffer needs to be larger than reported by sizeof(). So we > allocate a large buffer and cast it. Not sure if there is a trick to > get DEFINE_PER_CPU() to do that. So the array is variable length, but no larger than 4K - at least I hope because you're using a 4K buffer :) The neatest way to handle that is to make it a union, with the struct and a 4K char buffer. But we can do that as a cleanup later. > | > + memset(request_buffer, 0, 4096); > | > + memset(result_buffer, 0, 4096); > | > | Do we have to memset them? That's not going to speed things up. > > I agree about the speed, specially since we have a larger buffer. But we > are reusing the buffer for independent events and some fields need to be 0 > (hence the zalloc in the current code). Sure, so you could explicitly initialise those fields to zero. But that also can be another cleanup. I'll take this as it is. cheers
diff --git a/arch/powerpc/perf/hv-24x7.c b/arch/powerpc/perf/hv-24x7.c index dba3408..d073e06 100644 --- a/arch/powerpc/perf/hv-24x7.c +++ b/arch/powerpc/perf/hv-24x7.c @@ -217,11 +217,14 @@ static bool is_physical_domain(int domain) domain == HV_24X7_PERF_DOMAIN_PHYSICAL_CORE; } +DEFINE_PER_CPU(char, hv_24x7_reqb[4096]) __aligned(4096); +DEFINE_PER_CPU(char, hv_24x7_resb[4096]) __aligned(4096); + static unsigned long single_24x7_request(u8 domain, u32 offset, u16 ix, u16 lpar, u64 *res, bool success_expected) { - unsigned long ret = -ENOMEM; + unsigned long ret; /* * request_buffer and result_buffer are not required to be 4k aligned, @@ -243,13 +246,11 @@ static unsigned long single_24x7_request(u8 domain, u32 offset, u16 ix, BUILD_BUG_ON(sizeof(*request_buffer) > 4096); BUILD_BUG_ON(sizeof(*result_buffer) > 4096); - request_buffer = kmem_cache_zalloc(hv_page_cache, GFP_USER); - if (!request_buffer) - goto out; + request_buffer = (void *)get_cpu_var(hv_24x7_reqb); + result_buffer = (void *)get_cpu_var(hv_24x7_resb); - result_buffer = kmem_cache_zalloc(hv_page_cache, GFP_USER); - if (!result_buffer) - goto out_free_request_buffer; + memset(request_buffer, 0, 4096); + memset(result_buffer, 0, 4096); *request_buffer = (struct reqb) { .buf = { @@ -278,15 +279,11 @@ static unsigned long single_24x7_request(u8 domain, u32 offset, u16 ix, domain, offset, ix, lpar, ret, ret, result_buffer->buf.detailed_rc, result_buffer->buf.failing_request_ix); - goto out_free_result_buffer; + goto out; } *res = be64_to_cpu(result_buffer->result); -out_free_result_buffer: - kfree(result_buffer); -out_free_request_buffer: - kfree(request_buffer); out: return ret; }
Michael Ellerman [mpe@ellerman.id.au] wrote: | On Tue, 2014-12-09 at 23:06 -0800, Sukadev Bhattiprolu wrote: | > From 470c16c8955672103a9529c78dffbb239e9e27b8 Mon Sep 17 00:00:00 2001 | > From: Sukadev Bhattiprolu <sukadev@linux.vnet.ibm.com> | > Date: Tue, 9 Dec 2014 22:17:46 -0500 | > Subject: [PATCH 1/2] perf/poweprc/hv-24x7: Use per-cpu page buffer | > | > diff --git a/arch/powerpc/perf/hv-24x7.c b/arch/powerpc/perf/hv-24x7.c | > index dba3408..18e1f49 100644 | > --- a/arch/powerpc/perf/hv-24x7.c | > +++ b/arch/powerpc/perf/hv-24x7.c | > @@ -217,11 +217,14 @@ static bool is_physical_domain(int domain) | > domain == HV_24X7_PERF_DOMAIN_PHYSICAL_CORE; | > } | > | > +DEFINE_PER_CPU(char, hv_24x7_reqb[4096]); | > +DEFINE_PER_CPU(char, hv_24x7_resb[4096]); | | Do we need it to be 4K aligned also? I would guess so. Yes, fixed in the patch below. | | Rather than declaring these as char arrays and then casting below, can you pull | the struct definitions up and then declare the per cpu variables with the | proper type. Well, the structures, used for communication with HV, have variable length arrays, like: struct hv_24x7_request_buffer { ... struct hv_24x7_request requests[]; }; i.e the buffer needs to be larger than reported by sizeof(). So we allocate a large buffer and cast it. Not sure if there is a trick to get DEFINE_PER_CPU() to do that. We could add code to allocate pages per cpu during init, but that would mean more code, error handling etc. | | | > static unsigned long single_24x7_request(u8 domain, u32 offset, u16 ix, | > u16 lpar, u64 *res, | > bool success_expected) | > { | > - unsigned long ret = -ENOMEM; | > + unsigned long ret; | > | > /* | > * request_buffer and result_buffer are not required to be 4k aligned, | > @@ -243,13 +246,11 @@ static unsigned long single_24x7_request(u8 domain, u32 offset, u16 ix, | > BUILD_BUG_ON(sizeof(*request_buffer) > 4096); | > BUILD_BUG_ON(sizeof(*result_buffer) > 4096); | > | > - request_buffer = kmem_cache_zalloc(hv_page_cache, GFP_USER); | > - if (!request_buffer) | > - goto out; | > + request_buffer = (void *)get_cpu_var(hv_24x7_reqb); | > + result_buffer = (void *)get_cpu_var(hv_24x7_resb); | > | > - result_buffer = kmem_cache_zalloc(hv_page_cache, GFP_USER); | > - if (!result_buffer) | > - goto out_free_request_buffer; | > + memset(request_buffer, 0, 4096); | > + memset(result_buffer, 0, 4096); | | Do we have to memset them? That's not going to speed things up. I agree about the speed, specially since we have a larger buffer. But we are reusing the buffer for independent events and some fields need to be 0 (hence the zalloc in the current code). --- From 545ad0771ef8f22085ed8bb486f0a81b3afc5757 Mon Sep 17 00:00:00 2001 From: Sukadev Bhattiprolu <sukadev@linux.vnet.ibm.com> Date: Tue, 9 Dec 2014 22:17:46 -0500 Subject: [PATCH 1/2] powerpc/perf/hv-24x7: Use per-cpu page buffer The 24x7 counters are continuously running and not updated on an interrupt. So we record the event counts when stopping the event or deleting it. But to "read" a single counter in 24x7, we allocate a page and pass it into the hypervisor (The HV returns the page full of counters from which we extract the specific counter for this event). We allocate a page using GFP_USER and when deleting the event, we end up with the following warning because we are blocking in interrupt context. [ 698.641709] BUG: scheduling while atomic: swapper/0/0/0x10010000 We could use GFP_ATOMIC but that could result in failures. Pre-allocate a buffer so we don't have to allocate in interrupt context. Further as Michael Ellerman suggested, use Per-CPU buffer so we only need to allocate once per CPU. Signed-off-by: Sukadev Bhattiprolu <sukadev@linux.vnet.ibm.com> --- arch/powerpc/perf/hv-24x7.c | 21 +++++++++------------ 1 file changed, 9 insertions(+), 12 deletions(-) Changelog[v2]: [Michael Ellerman] Ensure buffer is aligned to 4K.