Date: Thu, 9 Oct 2014 17:17:38 -0400 From: John Baldwin <jhb@freebsd.org> To: Konstantin Belousov <kostikbel@gmail.com> Cc: svn-src-head@freebsd.org, Adrian Chadd <adrian@freebsd.org>, src-committers@freebsd.org, svn-src-all@freebsd.org Subject: Re: svn commit: r272800 - head/sys/x86/acpica Message-ID: <201410091717.38641.jhb@freebsd.org> In-Reply-To: <20141009182310.GL2153@kib.kiev.ua> References: <201410090534.s995YTUx057314@svn.freebsd.org> <20141009182310.GL2153@kib.kiev.ua>
next in thread | previous in thread | raw e-mail | index | archive | help
On Thursday, October 09, 2014 2:23:10 pm Konstantin Belousov wrote: > On Thu, Oct 09, 2014 at 05:34:29AM +0000, Adrian Chadd wrote: > > Author: adrian > > Date: Thu Oct 9 05:34:28 2014 > > New Revision: 272800 > > URL: https://svnweb.freebsd.org/changeset/base/272800 > > > > Log: > > Missing from previous commit - keep the VM domain -> PXM mapping > > array and use it to map PXM -> VM domain when needed. > > > > Differential Revision: D906 > > Reviewed by: jhb > > > > Modified: > > head/sys/x86/acpica/srat.c > > > > Modified: head/sys/x86/acpica/srat.c > > ============================================================================== > > --- head/sys/x86/acpica/srat.c Thu Oct 9 05:33:25 2014 (r272799) > > +++ head/sys/x86/acpica/srat.c Thu Oct 9 05:34:28 2014 (r272800) > > @@ -62,6 +62,8 @@ int num_mem; > > static ACPI_TABLE_SRAT *srat; > > static vm_paddr_t srat_physaddr; > > > > +static int vm_domains[VM_PHYSSEG_MAX]; > > + > > static void srat_walk_table(acpi_subtable_handler *handler, void *arg); > > > > /* > > @@ -247,7 +249,6 @@ check_phys_avail(void) > > static int > > renumber_domains(void) > > { > > - int domains[VM_PHYSSEG_MAX]; > > int i, j, slot; > > > > /* Enumerate all the domains. */ > > @@ -255,17 +256,17 @@ renumber_domains(void) > > for (i = 0; i < num_mem; i++) { > > /* See if this domain is already known. */ > > for (j = 0; j < vm_ndomains; j++) { > > - if (domains[j] >= mem_info[i].domain) > > + if (vm_domains[j] >= mem_info[i].domain) > > break; > > } > > - if (j < vm_ndomains && domains[j] == mem_info[i].domain) > > + if (j < vm_ndomains && vm_domains[j] == mem_info[i].domain) > > continue; > > > > /* Insert the new domain at slot 'j'. */ > > slot = j; > > for (j = vm_ndomains; j > slot; j--) > > - domains[j] = domains[j - 1]; > > - domains[slot] = mem_info[i].domain; > > + vm_domains[j] = vm_domains[j - 1]; > > + vm_domains[slot] = mem_info[i].domain; > > vm_ndomains++; > > if (vm_ndomains > MAXMEMDOM) { > > vm_ndomains = 1; > > @@ -280,15 +281,15 @@ renumber_domains(void) > > * If the domain is already the right value, no need > > * to renumber. > > */ > > - if (domains[i] == i) > > + if (vm_domains[i] == i) > > continue; > > > > /* Walk the cpu[] and mem_info[] arrays to renumber. */ > > for (j = 0; j < num_mem; j++) > > - if (mem_info[j].domain == domains[i]) > > + if (mem_info[j].domain == vm_domains[i]) > > mem_info[j].domain = i; > > for (j = 0; j <= MAX_APIC_ID; j++) > > - if (cpus[j].enabled && cpus[j].domain == domains[i]) > > + if (cpus[j].enabled && cpus[j].domain == vm_domains[i]) > > cpus[j].domain = i; > > } > > KASSERT(vm_ndomains > 0, > > @@ -368,4 +369,23 @@ srat_set_cpus(void *dummy) > > } > > } > > SYSINIT(srat_set_cpus, SI_SUB_CPU, SI_ORDER_ANY, srat_set_cpus, NULL); > > + > > +/* > > + * Map a _PXM value to a VM domain ID. > > + * > > + * Returns the domain ID, or -1 if no domain ID was found. > > + */ > > +int > > +acpi_map_pxm_to_vm_domainid(int pxm) > > +{ > > + int i; > > + > > + for (i = 0; i < vm_ndomains; i++) { > > + if (vm_domains[i] == pxm) > > + return (i); > > + } > > + > > + return (-1); > > +} > > + > > #endif /* MAXMEMDOM > 1 */ > > I do not like it. Sorry for not looking at the web thing, I have very > little time. > > It never was an intention that one proximity domain reported by ACPI > was mapped to single VM domain. VM could split domains (in terms of > vm_domains) further for other reasons. Main motivation is that there > is 1:1 relations between domain/page queues/page queues locks/pagedaemons. > > I have patches in WIP stage which split firmware proximity domains > further, to decrease congestion on the page queue locks. I wrote about > this in the pgsql performance report. > > The short version is that there is/will be N:1 relation between VM domains > and proximity domains (which is reported by ACPI for devices). _PXM is also defined to be what SRAT reports for memory, and is what will be used to do NUMA-aware memory allocations. While the VM system may decide to split a given NUMA domain into multiple some-other-things, those some-other- things won't be a NUMA domain anymore. At that point, you will need to divorce them from 'domain' and use another term as the domain index into mem_info[] will still be needed so that NUMA allocations do the correct thing. -- John Baldwin
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?201410091717.38641.jhb>