[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <5343d12a-630c-4d54-91f1-7a7d08326840@vivo.com>
Date: Thu, 14 Sep 2023 20:02:59 +0800
From: Bixuan Cui <cuibixuan@...o.com>
To: Jonathan Corbet <corbet@....net>,
Chuyi Zhou <zhouchuyi@...edance.com>, hannes@...xchg.org,
mhocko@...nel.org, roman.gushchin@...ux.dev, ast@...nel.org,
daniel@...earbox.net, andrii@...nel.org, muchun.song@...ux.dev
Cc: bpf@...r.kernel.org, linux-kernel@...r.kernel.org,
wuyun.abel@...edance.com, robin.lu@...edance.com
Subject: Re: [RFC PATCH v2 2/5] mm: Add policy_name to identify OOM policies
在 2023/8/15 4:51, Jonathan Corbet 写道:
>> /**
>> * dump_tasks - dump current memory state of all system tasks
>> * @oc: pointer to struct oom_control
>> @@ -484,8 +513,8 @@ static void dump_oom_summary(struct oom_control *oc, struct task_struct *victim)
>>
>> static void dump_header(struct oom_control *oc, struct task_struct *p)
>> {
>> - pr_warn("%s invoked oom-killer: gfp_mask=%#x(%pGg), order=%d, oom_score_adj=%hd\n",
>> - current->comm, oc->gfp_mask, &oc->gfp_mask, oc->order,
>> + pr_warn("%s invoked oom-killer: gfp_mask=%#x(%pGg), order=%d, policy_name=%s, oom_score_adj=%hd\n",
>> + current->comm, oc->gfp_mask, &oc->gfp_mask, oc->order, oc->policy_name,
> ...and if the policy name is unterminated, this print will run off the
> end of the structure.
>
> Am I missing something here?
Perhaps it is inaccurate to use policy name in this way. For example,
some one use BPF_PROG(bpf_oom_evaluate_task, ...) but do not set the
policy name through bpf_set_policy_name. In this way, the result is
still policy name=default, which ultimately leads to error print in the
dump_header.
I think a better way:
+static const char *const policy_select[] = {
+ "OOM_DEFAULT";
+ "BPF_ABORT",
+ "BPF_NEXT",
+ "BPF_SELECT",
+};
struct oom_control {
/* Used to print the constraint info. */
enum oom_constraint constraint;
+
+ /* Used to report the policy select. */
+ int policy_select;
};
static int oom_evaluate_task(struct task_struct *task, void *arg)
{
...
+ switch (bpf_oom_evaluate_task(task, oc)) {
+ case BPF_EVAL_ABORT:
+ oc->policy_select = BPF_EVAL_ABORT;
+ goto abort; /* abort search process */
+ case BPF_EVAL_NEXT:
+ oc->policy_select = BPF_EVAL_NEXT;
+ goto next; /* ignore the task */
+ case BPF_EVAL_SELECT:
+ oc->policy_select = BPF_EVAL_SELECT;
+ goto select; /* select the task */
+ default:
+ break; /* No BPF policy */
+ }
static void dump_header(struct oom_control *oc, struct task_struct *p)
{
- pr_warn("%s invoked oom-killer: gfp_mask=%#x(%pGg), order=%d,
oom_score_adj=%hd\n",
- current->comm, oc->gfp_mask, &oc->gfp_mask, oc->order,
+ pr_warn("%s invoked oom-killer: gfp_mask=%#x(%pGg), order=%d,
policy_name=%s, oom_score_adj=%hd\n",
+ current->comm, oc->gfp_mask, &oc->gfp_mask, oc->order,
policy_select[oc->policy_select],
current->signal->oom_score_adj);
And all definitions of oc should be added
struct oom_control oc = {
.select = NO_BPF_POLICY,
}
Delete set_oom_policy_name, it makes no sense for users to set policy
names. :-)
Thanks
Bixuan Cui
Powered by blists - more mailing lists