[lkml]   [2021]   [Feb]   [4]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
Patch in this message
Subject[PATCH RESEND v5] proc: Allow pid_revalidate() during LOOKUP_RCU
The pid_revalidate() function drops from RCU into REF lookup mode. When
many threads are resolving paths within /proc in parallel, this can
result in heavy spinlock contention on d_lockref as each thread tries to
grab a reference to the /proc dentry (and drop it shortly thereafter).

Investigation indicates that it is not necessary to drop RCU in
pid_revalidate(), as no RCU data is modified and the function never
sleeps. So, remove the LOOKUP_RCU check.

Signed-off-by: Stephen Brennan <>

When running running ~100 parallel instances of "TZ=/etc/localtime ps -fe
>/dev/null" on a 100CPU machine, the %sys utilization reaches 90%, and perf
shows the following code path as being responsible for heavy contention on
the d_lockref spinlock:

pid_revalidate() // returns -ECHILD
lockref_get_not_dead(&nd->path.dentry->d_lockref) <-- contention

By applying this patch, %sys utilization falls to around 60% under the same
workload. Although this particular workload is a bit contrived, we have
seen some monitoring scripts which produced similarly high %sys time due to
this contention.

As a result this patch, several procfs methods which were only called in
ref-walk mode could now be called from RCU mode. To ensure that this patch
is safe, I audited all the inode get_link and permission() implementations,
as well as dentry d_revalidate() implementations, in fs/proc. These methods
are called in the following ways:

* get_link() receives a NULL dentry pointer when called in RCU mode.
* permission() receives MAY_NOT_BLOCK in the mode parameter when called
from RCU.
* d_revalidate() receives LOOKUP_RCU in flags.

There were generally three groups I found. Group (1) are functions which
contain a check at the top of the function and return -ECHILD, and so
appear to be trivially RCU safe (although this is by dropping out of RCU
completely). Group (2) are functions which have no explicit check, but
on my audit, I was confident that there were no sleeping function calls,
and thus were RCU safe as is. However, I would appreciate any additional
review if possible. Group (3) are functions which might be be unsafe for some
reason or another.

Group (1):

Group (2):

Group (3):
pid_revalidate() -- addressed by my patch
proc_pid_permission() -- addressed by commits by Al
proc_map_files_get_link() -- calls capable() which could audit

I believe proc_map_files_get_link() is safe despite calling into the audit
framework, however I'm not confident and so I did not include it in Group 2.
proc_pid_permission() calls into the audit code, and is not safe with
these issues. This patch is tested and stable on the workload described
at the beginning of this cover letter, on a system with selinux enabled.

[1]: 23d8f5b684fc ("make dump_common_audit_data() safe to be called from
RCU pathwalk") and 2 previous

Changes in v5:
- Al's commits are now in linux-next, resolving proc_pid_permission() issue.
- Add NULL check after d_inode_rcu(dentry), because inode may become NULL if
we do not hold a reference.
Changes in v4:
- Simplify by unconditionally calling pid_update_inode() from pid_revalidate,
and removing the LOOKUP_RCU check.
Changes in v3:
- Rather than call pid_update_inode() with flags, create
proc_inode_needs_update() to determine whether the call can be skipped.
- Restore the call to the security hook (see next patch).
Changes in v2:
- Remove get_pid_task_rcu_user() and get_proc_task_rcu(), since they were
- Remove the call to security_task_to_inode().

fs/proc/base.c | 18 ++++++++++--------
1 file changed, 10 insertions(+), 8 deletions(-)

diff --git a/fs/proc/base.c b/fs/proc/base.c
index ebea9501afb8..3e105bd05801 100644
--- a/fs/proc/base.c
+++ b/fs/proc/base.c
@@ -1830,19 +1830,21 @@ static int pid_revalidate(struct dentry *dentry, unsigned int flags)
struct inode *inode;
struct task_struct *task;
+ int ret = 0;

- if (flags & LOOKUP_RCU)
- return -ECHILD;
- inode = d_inode(dentry);
- task = get_proc_task(inode);
+ rcu_read_lock();
+ inode = d_inode_rcu(dentry);
+ if (!inode)
+ goto out;
+ task = pid_task(proc_pid(inode), PIDTYPE_PID);

if (task) {
pid_update_inode(task, inode);
- put_task_struct(task);
- return 1;
+ ret = 1;
- return 0;
+ rcu_read_unlock();
+ return ret;

static inline bool proc_inode_is_dead(struct inode *inode)
 \ /
  Last update: 2021-02-04 18:21    [W:40.441 / U:0.148 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site