| Message ID | 20180323230849.3754-1-ross.zwisler@linux.intel.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Fri, Mar 23, 2018 at 4:08 PM, Ross Zwisler <ross.zwisler@linux.intel.com> wrote: > For systems that don't support NUMA, numactl gives a loud and fatal error: > > # numactl -N 0 ls > numactl: This system does not support NUMA policy > > Follow this model in ndctl for NUMA based filtering: > > # ./ndctl/ndctl list --numa-node=0 > Error: This system does not support NUMA > > This is done instead of just quietly filtering out all dimms, regions and > namespaces because the NUMA node they were trying to match didn't exist in > the system. > > libnuma tests whether NUMA is enabled via the get_mempolicy() syscall, > passing in all NULLs and 0s for arguments to always get the default policy. > See numa_available() in numa(3) and in the numactl source. > > ndctl checks sysfs for the existence of the /sys/devices/system/node > directory to avoid a dependency on libnuma. If we had a dependency on > libnuma we would have to choose whether this was fulfilled or not at > compile time, which would potentially mean that we could be on a > NUMA-enabled kernel but with an ndctl where NUMA support was disabled. > It's better to always have NUMA support in ndctl and only depend on the > kernel config. > > I've inspected the code for both get_mempolicy() and the code that creates > the /sys/devices/system/node directory, and they both seem to completely > rely on CONFIG_NUMA being defined. If CONFIG_NUMA is set, get_mempolicy() > will always be able to return a default policy and /sys/devices/system/node > will always exist. Otherwise, both checks will always fail. So, numactl > and ndctl should always agree on whether NUMA is supported on a given > system. > > Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> > Suggested-by: Dan Williams <dan.j.williams@intel.com> > --- > > v3: Changed back to checking /sys/devices/system/node instead of using > libnuma, and added more info to the changelog. Looks good, applied.
diff --git a/util/filter.c b/util/filter.c index 291d7ed..6ab391a 100644 --- a/util/filter.c +++ b/util/filter.c @@ -14,7 +14,10 @@ #include <string.h> #include <stdlib.h> #include <limits.h> +#include <unistd.h> +#include <sys/stat.h> #include <util/util.h> +#include <sys/types.h> #include <ndctl/ndctl.h> #include <util/filter.h> #include <ndctl/libndctl.h> @@ -328,6 +331,13 @@ int util_filter_walk(struct ndctl_ctx *ctx, struct util_filter_ctx *fctx, } if (param->numa_node && strcmp(param->numa_node, "all") != 0) { + struct stat st; + + if (stat("/sys/devices/system/node", &st) != 0) { + error("This system does not support NUMA"); + return -EINVAL; + } + numa_node = strtol(param->numa_node, &end, 0); if (end == param->numa_node || end[0]) { error("invalid numa_node: '%s'\n", param->numa_node);
For systems that don't support NUMA, numactl gives a loud and fatal error: # numactl -N 0 ls numactl: This system does not support NUMA policy Follow this model in ndctl for NUMA based filtering: # ./ndctl/ndctl list --numa-node=0 Error: This system does not support NUMA This is done instead of just quietly filtering out all dimms, regions and namespaces because the NUMA node they were trying to match didn't exist in the system. libnuma tests whether NUMA is enabled via the get_mempolicy() syscall, passing in all NULLs and 0s for arguments to always get the default policy. See numa_available() in numa(3) and in the numactl source. ndctl checks sysfs for the existence of the /sys/devices/system/node directory to avoid a dependency on libnuma. If we had a dependency on libnuma we would have to choose whether this was fulfilled or not at compile time, which would potentially mean that we could be on a NUMA-enabled kernel but with an ndctl where NUMA support was disabled. It's better to always have NUMA support in ndctl and only depend on the kernel config. I've inspected the code for both get_mempolicy() and the code that creates the /sys/devices/system/node directory, and they both seem to completely rely on CONFIG_NUMA being defined. If CONFIG_NUMA is set, get_mempolicy() will always be able to return a default policy and /sys/devices/system/node will always exist. Otherwise, both checks will always fail. So, numactl and ndctl should always agree on whether NUMA is supported on a given system. Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com> Suggested-by: Dan Williams <dan.j.williams@intel.com> --- v3: Changed back to checking /sys/devices/system/node instead of using libnuma, and added more info to the changelog. --- util/filter.c | 10 ++++++++++ 1 file changed, 10 insertions(+)