dcache: Translating dentry into pathname without taking rename_lock
When running the AIM7's short workload, Linus' lockref patch eliminated most of the spinlock contention. However, there were still some left: 8.46% reaim [kernel.kallsyms] [k] _raw_spin_lock |--42.21%-- d_path | proc_pid_readlink | SyS_readlinkat | SyS_readlink | system_call | __GI___readlink | |--40.97%-- sys_getcwd | system_call | __getcwd The big one here is the rename_lock (seqlock) contention in d_path() and the getcwd system call. This patch will eliminate the need to take the rename_lock while translating dentries into the full pathnames. The need to take the rename_lock is to make sure that no rename operation can be ongoing while the translation is in progress. However, only one thread can take the rename_lock thus blocking all the other threads that need it even though the translation process won't make any change to the dentries. This patch will replace the writer's write_seqlock/write_sequnlock sequence of the rename_lock of the callers of the prepend_path() and __dentry_path() functions with the reader's read_seqbegin/read_seqretry sequence within these 2 functions. As a result, the code will have to retry if one or more rename operations had been performed. In addition, RCU read lock will be taken during the translation process to make sure that no dentries will go away. To prevent live-lock from happening, the code will switch back to take the rename_lock if read_seqretry() fails for three times. To further reduce spinlock contention, this patch does not take the dentry's d_lock when copying the filename from the dentries. Instead, it treats the name pointer and length as unreliable and just copy the string byte-by-byte over until it hits a null byte or the end of string as specified by the length. This should avoid stepping into invalid memory address. The error cases are left to be handled by the sequence number check. The following code re-factoring are also made: 1. Move prepend('/') into prepend_name() to remove one conditional check. 2. Move the global root check in prepend_path() back to the top of the while loop. With this patch, the _raw_spin_lock will now account for only 1.2% of the total CPU cycles for the short workload. This patch also has the effect of reducing the effect of running perf on its profile since the perf command itself can be a heavy user of the d_path() function depending on the complexity of the workload. When taking the perf profile of the high-systime workload, the amount of spinlock contention contributed by running perf without this patch was about 16%. With this patch, the spinlock contention caused by the running of perf will go away and we will have a more accurate perf profile. Signed-off-by: Waiman Long <Waiman.Long@hp.com> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Showing
Please register or sign in to comment