Message ID | 20090902060911.GA5728@in.ibm.com (mailing list archive) |
---|---|
State | Deferred |
Headers | show |
On Wed, 2 Sep 2009, Ankita Garg wrote: > Hi, > > Below is a patch to fix a couple of issues with fake numa node creation > on ppc: > > 1) Presently, fake nodes could be created such that real numa node > boundaries are not respected. So a node could have lmbs that belong to > different real nodes. > On x86_64, we can use numa=off to completely disable NUMA so that all memory and all cpus are mapped to a single node 0. That's an extreme example of the above and is totally permissible. > 2) The cpu association is broken. On a JS22 blade for example, which is > a 2-node numa machine, I get the following: > > # cat /proc/cmdline > root=/dev/sda6 numa=fake=2G,4G,,6G,8G,10G,12G,14G,16G > # cat /sys/devices/system/node/node0/cpulist > 0-3 > # cat /sys/devices/system/node/node1/cpulist > 4-7 > # cat /sys/devices/system/node/node4/cpulist > > # > This doesn't show what the true NUMA topology of the machine is, could you please post the output of $ cat /sys/devices/system/node/node*/cpulist $ cat /sys/devices/system/node/node*/distance $ ls -d /sys/devices/system/node/node*/cpu[0-8] from a normal boot without any numa=fake? > So, though the cpus 4-7 should have been associated with node4, they > still belong to node1. The patch works by recording a real numa node > boundary and incrementing the fake node count. At the same time, a > mapping is stored from the real numa node to the first fake node that > gets created on it. > If there are multiple fake nodes on a real physical node, all cpus in that node should appear in the cpulist for each fake node for which it has local distance.
Hi David, On Tue, Sep 01, 2009 at 11:37:05PM -0700, David Rientjes wrote: > On Wed, 2 Sep 2009, Ankita Garg wrote: > > > Hi, > > > > Below is a patch to fix a couple of issues with fake numa node creation > > on ppc: > > > > 1) Presently, fake nodes could be created such that real numa node > > boundaries are not respected. So a node could have lmbs that belong to > > different real nodes. > > > > On x86_64, we can use numa=off to completely disable NUMA so that all > memory and all cpus are mapped to a single node 0. That's an extreme > example of the above and is totally permissible. > > > 2) The cpu association is broken. On a JS22 blade for example, which is > > a 2-node numa machine, I get the following: > > > > # cat /proc/cmdline > > root=/dev/sda6 numa=fake=2G,4G,,6G,8G,10G,12G,14G,16G > > # cat /sys/devices/system/node/node0/cpulist > > 0-3 > > # cat /sys/devices/system/node/node1/cpulist > > 4-7 > > # cat /sys/devices/system/node/node4/cpulist > > > > # > > > > This doesn't show what the true NUMA topology of the machine is, could you > please post the output of > > $ cat /sys/devices/system/node/node*/cpulist > $ cat /sys/devices/system/node/node*/distance > $ ls -d /sys/devices/system/node/node*/cpu[0-8] > > from a normal boot without any numa=fake? > Heres the output as requested by you: # ls /sys/devices/system/node/ has_cpu has_normal_memory node0 node1 online possible # cat /sys/devices/system/node/node*/cpulist 0-3 4-7 # cat /sys/devices/system/node/node*/distance 10 20 20 10 # ls -d /sys/devices/system/node/node*/cpu[0-8] /sys/devices/system/node/node0/cpu0 /sys/devices/system/node/node0/cpu3 /sys/devices/system/node/node1/cpu6 /sys/devices/system/node/node0/cpu1 /sys/devices/system/node/node1/cpu4 /sys/devices/system/node/node1/cpu7 /sys/devices/system/node/node0/cpu2 /sys/devices/system/node/node1/cpu5 > > So, though the cpus 4-7 should have been associated with node4, they > > still belong to node1. The patch works by recording a real numa node > > boundary and incrementing the fake node count. At the same time, a > > mapping is stored from the real numa node to the first fake node that > > gets created on it. > > > > If there are multiple fake nodes on a real physical node, all cpus in that > node should appear in the cpulist for each fake node for which it has > local distance. Currently, the behavior of fake numa is not so on x86 as well? Below is a sample output from a single node x86 system booted with numa=fake=8: # cat node0/cpulist # cat node1/cpulist ... # cat node6/cpulist # cat node7/cpulist 0-7 Presently, just fixing the cpu association issue with ppc, as explained in my previous mail.
On Wed, 2 Sep 2009, Ankita Garg wrote: > Currently, the behavior of fake numa is not so on x86 as well? Below is > a sample output from a single node x86 system booted with numa=fake=8: > > # cat node0/cpulist > > # cat node1/cpulist > > ... > # cat node6/cpulist > > # cat node7/cpulist > 0-7 > > Presently, just fixing the cpu association issue with ppc, as explained > in my previous mail. > Right, I'm proposing an alternate mapping scheme (which we've used for years) for both platforms such that a cpu is bound (and is set in cpumask_of_node()) to each fake node with which it has physical affinity. That is the only way for zonelist ordering in node order, task migration from offlined cpus, correct sched domains, etc. I can propose a patchset for x86_64 to do exactly this if there aren't any objections and I hope you'll help do ppc.
On Thu, Sep 3, 2009 at 1:06 AM, David Rientjes<rientjes@google.com> wrote: > On Wed, 2 Sep 2009, Ankita Garg wrote: > >> Currently, the behavior of fake numa is not so on x86 as well? Below is >> a sample output from a single node x86 system booted with numa=fake=8: >> >> # cat node0/cpulist >> >> # cat node1/cpulist >> >> ... >> # cat node6/cpulist >> >> # cat node7/cpulist >> 0-7 >> >> Presently, just fixing the cpu association issue with ppc, as explained >> in my previous mail. >> > > Right, I'm proposing an alternate mapping scheme (which we've used for > years) for both platforms such that a cpu is bound (and is set in > cpumask_of_node()) to each fake node with which it has physical affinity. > That is the only way for zonelist ordering in node order, task migration > from offlined cpus, correct sched domains, etc. I can propose a patchset > for x86_64 to do exactly this if there aren't any objections and I hope > you'll help do ppc. Sounds interesting, I'd definitely be interested in seeing your proposal, but I would think of that as additional development on top of this patch Balbir Singh.
On Thu, 3 Sep 2009, Balbir Singh wrote: > > Right, I'm proposing an alternate mapping scheme (which we've used for > > years) for both platforms such that a cpu is bound (and is set in > > cpumask_of_node()) to each fake node with which it has physical affinity. > > That is the only way for zonelist ordering in node order, task migration > > from offlined cpus, correct sched domains, etc. I can propose a patchset > > for x86_64 to do exactly this if there aren't any objections and I hope > > you'll help do ppc. > > Sounds interesting, I'd definitely be interested in seeing your > proposal, but I would think of that as additional development on top > of this patch > Absolutely. I'm not familiar with numa=fake on ppc, but if cpus are being bound to nodes with which they don't have affinity, it definitely warrants a fix such as this (although the initial value for fake_enabled looks wrong and fake_numa_node_mapping[] can be __cpuinitdata). I'll cc you, Ben, and Ankita on the x86_64 patches. Thanks.
different real nodes. 2) The cpu association is broken. On a JS22 blade for example, which is a 2-node numa machine, I get the following: # cat /proc/cmdline root=/dev/sda6 numa=fake=2G,4G,,6G,8G,10G,12G,14G,16G # cat /sys/devices/system/node/node0/cpulist 0-3 # cat /sys/devices/system/node/node1/cpulist 4-7 # cat /sys/devices/system/node/node4/cpulist # So, though the cpus 4-7 should have been associated with node4, they still belong to node1. The patch works by recording a real numa node boundary and incrementing the fake node count. At the same time, a mapping is stored from the real numa node to the first fake node that gets created on it. Tested the patch with the following commandlines: numa=fake=2G,4G,6G,8G,10G,12G,14G,16G numa=fake=3G,6G,10G,16G numa=fake=4G numa=fake= For testing if the fake nodes respect the real node boundaries, I added some debug printks in the node creation path. Without the patch, for the commandline numa=fake=2G,4G,6G,8G,10G,12G,14G,16G, this is what I got: fake id: 1 nid: 0 fake id: 1 nid: 0 ... fake id: 2 nid: 0 fake id: 2 nid: 0 ... fake id: 2 nid: 0 created new fake_node with id 3 fake id: 3 nid: 0 fake id: 3 nid: 0 ... fake id: 3 nid: 0 fake id: 3 nid: 0 fake id: 3 nid: 1 fake id: 3 nid: 1 ... created new fake_node with id 4 fake id: 4 nid: 1 fake id: 4 nid: 1 ... and so on. So, fake node 3 encompasses real node 0 & 1. Also, # cat /sys/devices/system/node/node3/meminfo Node 0 MemTotal: 2097152 kB ... # # cat /sys/devices/system/node/node4/meminfo Node 0 MemTotal: 2097152 kB ... With the patch, I get: fake id: 1 nid: 0 fake id: 1 nid: 0 ... fake id: 2 nid: 0 fake id: 2 nid: 0 ... fake id: 2 nid: 0 created new fake_node with id 3 fake id: 3 nid: 0 fake id: 3 nid: 0 ... fake id: 3 nid: 0 fake id: 3 nid: 0 created new fake_node with id 4 fake id: 4 nid: 1 fake id: 4 nid: 1 ... and so on. With the patch, the fake node sizes are slightly different from that specified by the user. # cat /sys/devices/system/node/node3/meminfo Node 3 MemTotal: 1638400 kB ... # cat /sys/devices/system/node/node4/meminfo Node 4 MemTotal: 458752 kB ... CPU association was tested as mentioned in the previous mail: Without the patch, # cat /proc/cmdline root=/dev/sda6 numa=fake=2G,4G,,6G,8G,10G,12G,14G,16G # cat /sys/devices/system/node/node0/cpulist 0-3 # cat /sys/devices/system/node/node1/cpulist 4-7 # cat /sys/devices/system/node/node4/cpulist # With the patch, # cat /proc/cmdline root=/dev/sda6 numa=fake=2G,4G,,6G,8G,10G,12G,14G,16G # cat /sys/devices/system/node/node0/cpulist 0-3 # cat /sys/devices/system/node/node1/cpulist # cat /sys/devices/system/node/node4/cpulist 4-7 Signed-off-by: Ankita Garg <ankita@in.ibm.com> Reviewed-by: Balbir Singh <balbir@linux.vnet.ibm.com> Index: linux-2.6.31-rc5/arch/powerpc/mm/numa.c =================================================================== --- linux-2.6.31-rc5.orig/arch/powerpc/mm/numa.c +++ linux-2.6.31-rc5/arch/powerpc/mm/numa.c @@ -26,6 +26,13 @@ #include <asm/smp.h> static int numa_enabled = 1; +static int fake_enabled = 1; + +/* + * The array maps a real numa node to the first fake node that gets + * created on it + */ +int fake_numa_node_mapping[MAX_NUMNODES]; static char *cmdline __initdata; @@ -49,14 +56,29 @@ static int __cpuinit fake_numa_create_ne unsigned long long mem; char *p = cmdline; static unsigned int fake_nid; + static unsigned int prev_nid = 0; static unsigned long long curr_boundary; /* * Modify node id, iff we started creating NUMA nodes * We want to continue from where we left of the last time */ - if (fake_nid) + if (fake_nid) { + /* + * Moved over to the next real numa node, increment fake + * node number and store the mapping of the real node to + * the fake node + */ + if (prev_nid != *nid) { + fake_nid++; + fake_numa_node_mapping[*nid] = fake_nid; + prev_nid = *nid; + *nid = fake_nid; + return 0; + } *nid = fake_nid; + } + /* * In case there are no more arguments to parse, the * node_id should be the same as the last fake node id @@ -440,7 +462,7 @@ static int of_drconf_to_nid_single(struc */ static int __cpuinit numa_setup_cpu(unsigned long lcpu) { - int nid = 0; + int nid = 0, new_nid; struct device_node *cpu = of_get_cpu_node(lcpu, NULL); if (!cpu) { @@ -450,8 +472,15 @@ static int __cpuinit numa_setup_cpu(unsi nid = of_node_to_nid_single(cpu); + if (fake_enabled && nid) { + new_nid = fake_numa_node_mapping[nid]; + if (new_nid > 0) + nid = new_nid; + } + if (nid < 0 || !node_online(nid)) nid = any_online_node(NODE_MASK_ALL); + out: map_cpu_to_node(lcpu, nid); @@ -1005,8 +1034,12 @@ static int __init early_numa(char *p) numa_debug = 1; p = strstr(p, "fake="); - if (p) + if (p) { cmdline = p + strlen("fake="); + if (numa_enabled) { + fake_enabled = 1; + } + } return 0; }