On 03/07/2024 06:28, Huang, Ying wrote:
Tvrtko Ursulin <tursulin@xxxxxxxxxx> writes:From: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxxx> Since balancing mode was added in bda420b98505 ("numa balancing: migrate on fault among multiple bound nodes"), it was possible to set this mode but it wouldn't be shown in /proc/<pid>/numa_maps since there was no support for it in the mpol_to_str() helper. Furthermore, because the balancing mode sets the MPOL_F_MORON flag, it would be displayed as 'default' due a workaround introduced a few years earlier in 8790c71a18e5 ("mm/mempolicy.c: fix mempolicy printing in numa_maps"). To tidy this up we implement two changes: First we introduce a new internal flag MPOL_F_KERNEL and with it mark the kernel's internal default and fallback policies (for tasks and/or VMAs with no explicit policy set). By doing this we generalise the current special casing and replace the incorrect 'default' with the correct 'bind'. Secondly, we add a string representation and corresponding handling for MPOL_F_NUMA_BALANCING. We do this by adding a sparse mapping array of flags to names. With the sparseness being the downside, but with the advantage of generalising and removing the "policy" from flags display.Please split these 2 changes into 2 patches. Because we will need to back port the first one to -stable kernel.
Why two? AFAICT there wasn't a issue until bda420b98505, and to fix it all changes from this patch are needed.
End result: $ numactl -b -m 0-1,3 cat /proc/self/numa_maps 555559580000 bind=balancing:0-1,3 file=/usr/bin/cat mapped=3 active=0 N0=3 kernelpagesize_kB=16 ... v2: * Fully fix by introducing MPOL_F_KERNEL. Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxxx> Fixes: bda420b98505 ("numa balancing: migrate on fault among multiple bound nodes") References: 8790c71a18e5 ("mm/mempolicy.c: fix mempolicy printing in numa_maps") Cc: Huang Ying <ying.huang@xxxxxxxxx> Cc: Mel Gorman <mgorman@xxxxxxx> Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx> Cc: Ingo Molnar <mingo@xxxxxxxxxx> Cc: Rik van Riel <riel@xxxxxxxxxxx> Cc: Johannes Weiner <hannes@xxxxxxxxxxx> Cc: "Matthew Wilcox (Oracle)" <willy@xxxxxxxxxxxxx> Cc: Dave Hansen <dave.hansen@xxxxxxxxx> Cc: Andi Kleen <ak@xxxxxxxxxxxxxxx> Cc: Michal Hocko <mhocko@xxxxxxxx> Cc: David Rientjes <rientjes@xxxxxxxxxx> --- include/uapi/linux/mempolicy.h | 1 + mm/mempolicy.c | 44 ++++++++++++++++++++++++---------- 2 files changed, 32 insertions(+), 13 deletions(-) diff --git a/include/uapi/linux/mempolicy.h b/include/uapi/linux/mempolicy.h index 1f9bb10d1a47..bcf56ce9603b 100644 --- a/include/uapi/linux/mempolicy.h +++ b/include/uapi/linux/mempolicy.h @@ -64,6 +64,7 @@ enum { #define MPOL_F_SHARED (1 << 0) /* identify shared policies */ #define MPOL_F_MOF (1 << 3) /* this policy wants migrate on fault */ #define MPOL_F_MORON (1 << 4) /* Migrate On protnone Reference On Node */ +#define MPOL_F_KERNEL (1 << 5) /* Kernel's internal policy *//** These bit locations are exposed in the vm.zone_reclaim_mode sysctl diff --git a/mm/mempolicy.c b/mm/mempolicy.c index aec756ae5637..8ecc6d9f100a 100644 --- a/mm/mempolicy.c +++ b/mm/mempolicy.c @@ -134,6 +134,7 @@ enum zone_type policy_zone = 0; static struct mempolicy default_policy = { .refcnt = ATOMIC_INIT(1), /* never free it */ .mode = MPOL_LOCAL, + .flags = MPOL_F_KERNEL, };static struct mempolicy preferred_node_policy[MAX_NUMNODES];@@ -3095,7 +3096,7 @@ void __init numa_policy_init(void) preferred_node_policy[nid] = (struct mempolicy) { .refcnt = ATOMIC_INIT(1), .mode = MPOL_PREFERRED, - .flags = MPOL_F_MOF | MPOL_F_MORON, + .flags = MPOL_F_MOF | MPOL_F_MORON | MPOL_F_KERNEL, .nodes = nodemask_of_node(nid), }; } @@ -3150,6 +3151,12 @@ static const char * const policy_modes[] = [MPOL_PREFERRED_MANY] = "prefer (many)", };+static const char * const policy_flags[] = {+ [ilog2(MPOL_F_STATIC_NODES)] = "static", + [ilog2(MPOL_F_RELATIVE_NODES)] = "relative", + [ilog2(MPOL_F_NUMA_BALANCING)] = "balancing", +}; + #ifdef CONFIG_TMPFS /** * mpol_parse_str - parse string to mempolicy, for tmpfs mpol mount option. @@ -3293,17 +3300,18 @@ int mpol_parse_str(char *str, struct mempolicy **mpol) * @pol: pointer to mempolicy to be formatted * * Convert @pol into a string. If @buffer is too short, truncate the string. - * Recommend a @maxlen of at least 32 for the longest mode, "interleave", the - * longest flag, "relative", and to display at least a few node ids. + * Recommend a @maxlen of at least 42 for the longest mode, "weighted + * interleave", the longest flag, "balancing", and to display at least a few + * node ids. */ void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol) { char *p = buffer; nodemask_t nodes = NODE_MASK_NONE; unsigned short mode = MPOL_DEFAULT; - unsigned short flags = 0; + unsigned long flags = 0;- if (pol && pol != &default_policy && !(pol->flags & MPOL_F_MORON)) {+ if (!(pol->flags & MPOL_F_KERNEL)) {Can we avoid to introduce a new flag? Whether the following code work? if (pol && pol != &default_policy && !(pol->mode != MPOL_PREFERRED) && !(pol->flags & MPOL_F_MORON)) But I think that this is kind of fragile. A flag is better. But personally, I don't think MPOL_F_KERNEL is a good name, maybe MPOL_F_DEFAULT?
I thought along the same lines, but as you have also shown we need to exclude both default and preferred fallbacks so naming the flag default did not feel best. MPOL_F_INTERNAL? MPOL_F_FALLBACK? MPOL_F_SHOW_AS_DEFAULT? :))
What I dislike about the flag more is the fact internal flags are for some reason in the uapi headers. And presumably we cannot zap them.
But I don't think we can check for MPOL_PREFERRED since it can be a legitimate user set policy.
We could check for the address of preferred_node_policy[] members with a loop covering all possible nids? If that will be the consensus I am happy to change it. But flag feels more elegant and robust.
mode = pol->mode; flags = pol->flags; } @@ -3328,15 +3336,25 @@ void mpol_to_str(char *buffer, int maxlen, struct mempolicy *pol) p += snprintf(p, maxlen, "%s", policy_modes[mode]);if (flags & MPOL_MODE_FLAGS) {- p += snprintf(p, buffer + maxlen - p, "="); + unsigned int bit, cnt = 0;- /*- * Currently, the only defined flags are mutually exclusive - */ - if (flags & MPOL_F_STATIC_NODES) - p += snprintf(p, buffer + maxlen - p, "static"); - else if (flags & MPOL_F_RELATIVE_NODES) - p += snprintf(p, buffer + maxlen - p, "relative"); + for_each_set_bit(bit, &flags, ARRAY_SIZE(policy_flags)) { + if (bit <= ilog2(MPOL_F_KERNEL)) + continue; + + if (cnt == 0) + p += snprintf(p, buffer + maxlen - p, "="); + else + p += snprintf(p, buffer + maxlen - p, ","); + + if (WARN_ON_ONCE(!policy_flags[bit])) + p += snprintf(p, buffer + maxlen - p, "bit%u", + bit); + else + p += snprintf(p, buffer + maxlen - p, + policy_flags[bit]); + cnt++; + }Please refer to commit 2291990ab36b ("mempolicy: clean-up mpol-to-str() mempolicy formatting") for the original format.
That was in 2008 so long time ago and in the meantime there were no bars. The format in this patch tries to align with the input format and I think it manages, apart from deciding to print unknown flags as bit numbers (which is most probably an irrelevant difference). Why do you think the pre-2008 format is better?
Regards, Tvrtko
}if (!nodes_empty(nodes))-- Best Regards, Huang, Ying
