9:45 p.m.
From: Kazuhito Hagio <k-hagio-ab(a)nec.com>
Kernel commit b8fea7af0e40 ("sched/fair: Allocate cfs_tg_state with
percpu allocator") changed task_group.cfs_rq from a pointer array to a
per-cpu variable allocated by the per-cpu allocator.
Without the patch, the "runq -g" option fails with the following error
on the first time, and with a segmentation fault on the second time.
crash> runq -g
CPU 0
CURRENT: PID: 5687 TASK: ffff8b0bc175a2c0 COMMAND: "bash"
ROOT_TASK_GROUP: ffffffff93a55600 CFS_RQ: 0
runq: invalid kernel virtual address: 58 type: "curr"
crash> runq -g
Segmentation fault (core dumped)
Since there is no way to determine whether it is a per-cpu variable,
check wthether it is a pointer array or not.
Signed-off-by: Kazuhito Hagio <k-hagio-ab(a)nec.com>
---
Hi,
I could not find a way to determine it's a __percpu variable or not,
any better ideas?
defs.h | 2 ++
kernel.c | 14 ++++++++++++++
symbols.c | 27 +++++++++++++++++++++++++++
task.c | 30 ++++++++++++++++++++++--------
4 files changed, 65 insertions(+), 8 deletions(-)
diff --git a/defs.h b/defs.h
index a4f70b773cd7..caa4faabb73e 100644
--- a/defs.h
+++ b/defs.h
@@ -691,6 +691,7 @@ struct new_utsname {
#define KMOD_PAX (0x100ULL)
#define KMOD_MEMORY (0x200ULL)
#define IRQ_DESC_TREE_MAPLE (0x400ULL)
+#define PER_CPU_CFS_RQ (0x800ULL)
#define XEN() (kt->flags & ARCH_XEN)
#define OPENVZ() (kt->flags & ARCH_OPENVZ)
@@ -5884,6 +5885,7 @@ void parse_for_member_extended(struct datatype_member *, ulong);
void add_to_downsized(char *);
int is_downsized(char *);
int is_string(char *, char *);
+int is_ptrptr(char *, char *);
struct syment *symbol_complete_match(const char *, struct syment *);
/*
diff --git a/kernel.c b/kernel.c
index eb9754c5e082..e53038d5d8db 100644
--- a/kernel.c
+++ b/kernel.c
@@ -400,6 +400,18 @@ kernel_init()
MEMBER_OFFSET_INIT(task_group_rt_rq, "task_group", "rt_rq");
MEMBER_OFFSET_INIT(task_group_parent, "task_group", "parent");
+ /*
+ * task_group.cfs_rq was changed from a pointer array to a per-cpu
+ * variable at Linux 7.2 (b8fea7af0e40). Since there is no way to
+ * determine it, we check whether it is a pointer array or not.
+ * - struct cfs_rq **cfs_rq;
+ * + struct cfs_rq __percpu *cfs_rq;
+ */
+ if (VALID_MEMBER(task_group_cfs_rq)) {
+ if (!is_ptrptr("task_group", "cfs_rq"))
+ kt->flags2 |= PER_CPU_CFS_RQ;
+ }
+
/*
* In 2.4, smp_send_stop() sets smp_num_cpus back to 1
* in some, but not all, architectures. So if a count
@@ -6362,6 +6374,8 @@ dump_kernel_table(int verbose)
fprintf(fp, "%sKMOD_PAX", others++ ? "|" : "");
if (kt->flags2 & KMOD_MEMORY)
fprintf(fp, "%sKMOD_MEMORY", others++ ? "|" : "");
+ if (kt->flags2 & PER_CPU_CFS_RQ)
+ fprintf(fp, "%sPER_CPU_CFS_RQ", others++ ? "|" : "");
fprintf(fp, ")\n");
fprintf(fp, " stext: %lx\n", kt->stext);
diff --git a/symbols.c b/symbols.c
index 78e400ba3756..03511c8cbe8c 100644
--- a/symbols.c
+++ b/symbols.c
@@ -7933,6 +7933,33 @@ is_string(char *structure, char *member)
return retval;
}
+int
+is_ptrptr(char *structure, char *member)
+{
+ int retval;
+ char *t;
+ char buf[BUFSIZE];
+
+ retval = FALSE;
+ open_tmpfile();
+ whatis_datatype(structure, STRUCT_REQUEST, pc->tmpfile);
+ rewind(pc->tmpfile);
+ while (fgets(buf, BUFSIZE, pc->tmpfile)) {
+ if (!(t = strstr(buf, "**")))
+ continue;
+ t += 2;
+ if (t != strstr(t, member))
+ continue;
+ t += strlen(member);
+ if (*t == ';') {
+ retval = TRUE;
+ break;
+ }
+ }
+ close_tmpfile();
+
+ return retval;
+}
/*
* Generic function for dumping data structure declarations, with a small
diff --git a/task.c b/task.c
index d7f3c129a709..5920d6499687 100644
--- a/task.c
+++ b/task.c
@@ -9464,8 +9464,12 @@ print_parent_task_group_fair(void *t, int cpu)
readmem(tgi->task_group + OFFSET(task_group_cfs_rq),
KVADDR, &cfs_rq_c, sizeof(ulong),
"task_group cfs_rq", FAULT_ON_ERROR);
- readmem(cfs_rq_c + cpu * sizeof(ulong), KVADDR, &cfs_rq_p,
- sizeof(ulong), "task_group cfs_rq", FAULT_ON_ERROR);
+
+ if (kt->flags2 & PER_CPU_CFS_RQ)
+ cfs_rq_p = cfs_rq_c + kt->__per_cpu_offset[cpu];
+ else
+ readmem(cfs_rq_c + cpu * sizeof(ulong), KVADDR, &cfs_rq_p,
+ sizeof(ulong), "task_group cfs_rq", FAULT_ON_ERROR);
print_group_header_fair(tgi->depth, cfs_rq_p, tgi);
tgi->use = 0;
@@ -9491,8 +9495,13 @@ dump_tasks_in_lower_dequeued_cfs_rq(int depth, ulong cfs_rq, int
cpu,
readmem(tgi_array[i]->task_group + OFFSET(task_group_cfs_rq),
KVADDR, &cfs_rq_c, sizeof(ulong), "task_group cfs_rq",
FAULT_ON_ERROR);
- readmem(cfs_rq_c + cpu * sizeof(ulong), KVADDR, &cfs_rq_p,
- sizeof(ulong), "task_group cfs_rq", FAULT_ON_ERROR);
+
+ if (kt->flags2 & PER_CPU_CFS_RQ)
+ cfs_rq_p = cfs_rq_c + kt->__per_cpu_offset[cpu];
+ else
+ readmem(cfs_rq_c + cpu * sizeof(ulong), KVADDR, &cfs_rq_p,
+ sizeof(ulong), "task_group cfs_rq", FAULT_ON_ERROR);
+
if (cfs_rq == cfs_rq_p)
continue;
@@ -10355,10 +10364,15 @@ dump_tasks_by_task_group(void)
readmem(rt_rq + cpu * sizeof(ulong), KVADDR,
&rt_rq_p, sizeof(ulong), "task_group rt_rq",
FAULT_ON_ERROR);
- if (cfs_rq)
- readmem(cfs_rq + cpu * sizeof(ulong), KVADDR,
- &cfs_rq_p, sizeof(ulong), "task_group cfs_rq",
- FAULT_ON_ERROR);
+ if (cfs_rq) {
+ if (kt->flags2 & PER_CPU_CFS_RQ)
+ cfs_rq_p = cfs_rq + kt->__per_cpu_offset[cpu];
+ else
+ readmem(cfs_rq + cpu * sizeof(ulong), KVADDR,
+ &cfs_rq_p, sizeof(ulong),
+ "task_group cfs_rq", FAULT_ON_ERROR);
+ }
+
fprintf(fp, "%sCPU %d", displayed++ ? "\n" : "", cpu);
if (hide_offline_cpu(cpu)) {
--
2.31.1