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 <[email protected]>
---
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:
walk_component()
lookup_fast()
d_revalidate()
pid_revalidate() // returns -ECHILD
unlazy_child()
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):
proc_ns_get_link()
proc_pid_get_link()
map_files_d_revalidate()
proc_misc_d_revalidate()
tid_fd_revalidate()
Group (2):
proc_get_link()
proc_self_get_link()
proc_thread_self_get_link()
proc_fd_permission()
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
LSM_AUDIT_DATA_DENTRY or LSM_AUDIT_DATA_INODE. Al's commits[1] address
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
unnecessary.
- 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;
+out:
+ rcu_read_unlock();
+ return ret;
}
static inline bool proc_inode_is_dead(struct inode *inode)
--
2.27.0
Stephen Brennan <[email protected]> writes:
> 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 <[email protected]>
> ---
Hello all,
I wanted to bring this patch to the surface again in case anyone has an
opportunity to review the changes.
Thank you,
Stephen
>
> 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:
>
> walk_component()
> lookup_fast()
> d_revalidate()
> pid_revalidate() // returns -ECHILD
> unlazy_child()
> 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):
> proc_ns_get_link()
> proc_pid_get_link()
> map_files_d_revalidate()
> proc_misc_d_revalidate()
> tid_fd_revalidate()
>
> Group (2):
> proc_get_link()
> proc_self_get_link()
> proc_thread_self_get_link()
> proc_fd_permission()
>
> 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
> LSM_AUDIT_DATA_DENTRY or LSM_AUDIT_DATA_INODE. Al's commits[1] address
> 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
> unnecessary.
> - 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;
> +out:
> + rcu_read_unlock();
> + return ret;
> }
>
> static inline bool proc_inode_is_dead(struct inode *inode)
> --
> 2.27.0