idmap parallelism improvement patch
Subsystems (2)
Hi,
Has this patch about idmap parallelism improvement been scheduled for a coming release?
Cheers.
Message: 8
Date: Wed, 04 Apr 2007 13:10:33 -0700
From: Frank Filz <[email protected]>
Subject: [NFS] [PATCH] [RESEND] Improve idmap parallelism
To: Trond Myklebust <[email protected]>
Cc: NFS List <[email protected]>
Message-ID: <1175717433.3531.13.camel@dyn9047022153>
Content-Type: text/plain
Not sure if this got lost in the noise, resending...
Quoted 13 lines · Trond Myklebust
Quoted 13 lines · Trond Myklebust
Quoted 12 lines
Quoted 12 lines · Frank Filz
Quoted 3 lines
Quoted 3 lines
Resend: I sent this awhile ago, but it may have been missed with the excitement of Connectathon, weddings, and honeymoons. I've verified it compiles against 2.6.21 (I originally tested it on 2.6.20).
You forgot the Linux Storage and Filesystem Workshop and FAST. It has been an exciting life in the past 2 months.
Quoted 8 lines
Quoted 6 lines
Quoted 1 line
Quoted 1 line
At first glance it looks OK. I'll take it for a spin...
Hmm... Apart from having been mangled by your mailer, I immediately started getting major complaints about scheduling under a spin lock.
My apologies on the formatting.
Quoted 6 lines
Quoted 4 lines
Looking more closely, it seems you are holding the read lock while calling
rpc_queue_upcall(), which again will result in a call toidmap_pipe_upcall(). That won't work sincecopy_to_user()is allowed to sleep.
Ah, oops. I have rewritten the code using rwsem to avoid this problem. This may not have caused a problem in my testing because the only contention should be the write, which shouldn't be attempted until after the copy_to_user is done (given the serialization of upcalls). Still, no excuse for faulty coding.
Quoted 3 lines
Quoted 1 line
In addition, there were several leakages of the read lock.
Fixed those.
I ran the same testing overnight on this updated patch.
Here is the re-submitted patch:
This patch improves idmap parallelism by reducing the serialization of idmap lookups. Currently, each client can only process one idmap lookup at a time, potentially delaying other lookups that might be satisfied by the in kernel cache while waiting for a user space lookup.
The existing code uses two mutexes, but one of them is held the entire time of the lookup.
The biggest change this patch makes is to re-order lock use so that one lock serializes multiple user space lookups (for the same nfs_client). This lock is only held when the kernel cache lookup fails and a user space lookup must occur. The other lock is used to protect the in kernel cache (a pair of hash tables). This lock is only held when the internal cache is being accessed. Further, since most accesses are lookups, this second lock is changed to a read-write lock.
After acquiring the mutex, a second cache lookup is made just in case a user space lookup had been in progress for this id.
I tested this using fsstress on an SMP machine. While testing, I put in some metering code which showed as many as 1000 cache lookups satisfied while an upcall was in progress, and noted occasional lookups for the id that an upcall was in progress for.
This patch was modified from an initial patch by Usha Ketineni.
fs/nfs/idmap.c +4 -3
@@ -36,6 +36,7 @@#include <linux/module.h> #include <linux/mutex.h> +#include <linux/rwsem.h> #include <linux/init.h> #include <linux/types.h> #include <linux/slab.h>
@@ -86,9 +87,9 @@ struct idmap_hashtable {struct idmap {struct dentry *idmap_dentry;wait_queue_head_t idmap_wq; -struct idmap_msg idmap_im; +struct idmap_msg idmap_im;/* protected by mutex */ struct mutex idmap_lock;/* Serializes upcalls */ -struct mutex idmap_im_lock;/* Protects the hashtable */ +struct rw_semaphore idmap_im_lock;/* Protects the hashtable */ struct idmap_hashtable idmap_user_hash;struct idmap_hashtable idmap_group_hash; };
@@ -127,7 +128,7 @@ nfs_idmap_new(struct nfs_client *clp) }mutex_init (&idmap->idmap_lock); -mutex_init (&idmap->idmap_im_lock); +init_rwsem (&idmap->idmap_im_lock);init_waitqueue_head (&idmap->idmap_wq); idmap->idmap_user_hash.h_type =IDMAP_TYPE_USER ; idmap->idmap_group_hash.h_type =IDMAP_TYPE_GROUP ;
@@ -243,14 +244,28 @@ nfs_idmap_id(struct idmap *idmap, struct
idmap_hashtable *h,
if (namelen >= IDMAP_NAMESZ)
return -EINVAL;
+ down_read(&idmap->idmap_im_lock);
+ he = idmap_lookup_name(h, name, namelen);
+ if (he != NULL) {
+ *id = he->ih_id;
+ up_read(&idmap->idmap_im_lock);
+ return 0;
+ }
+ up_read(&idmap->idmap_im_lock);
+
mutex_lock(&idmap->idmap_lock);
- mutex_lock(&idmap->idmap_im_lock);
+ /* Attempt lookup again in case we blocked
+ * because another attempt on this name
+ * was in progress.
+ */
+ down_read(&idmap->idmap_im_lock);
he = idmap_lookup_name(h, name, namelen);
if (he != NULL) {
*id = he->ih_id;
- ret = 0;
- goto out;
+ up_read(&idmap->idmap_im_lock);
+ mutex_unlock(&idmap->idmap_lock);
+ return 0;
}
memset(im, 0, sizeof(*im));
@@ -266,15 +281,15 @@ nfs_idmap_id(struct idmap *idmap, struct idmap_hashtable *h,
add_wait_queue(&idmap->idmap_wq, &wq);
if (rpc_queue_upcall(idmap->idmap_dentry->d_inode, &msg) < 0) {
remove_wait_queue(&idmap->idmap_wq, &wq);
+ up_read(&idmap->idmap_im_lock);
goto out;
}
set_current_state(TASK_UNINTERRUPTIBLE);
- mutex_unlock(&idmap->idmap_im_lock);
+ up_read(&idmap->idmap_im_lock);
schedule();
current->state = TASK_RUNNING;
remove_wait_queue(&idmap->idmap_wq, &wq);
- mutex_lock(&idmap->idmap_im_lock);
if (im->im_status & IDMAP_STATUS_SUCCESS) {
*id = im->im_id;
@@ -283,7 +298,6 @@ nfs_idmap_id(struct idmap *idmap, struct idmap_hashtable *h,
out:
memset(im, 0, sizeof(*im));
- mutex_unlock(&idmap->idmap_im_lock);
mutex_unlock(&idmap->idmap_lock);
return (ret);
}
@@ -304,14 +318,30 @@ nfs_idmap_name(struct idmap *idmap, struct idmap_hashtable *h,
im = &idmap->idmap_im;
+ down_read(&idmap->idmap_im_lock);
+ he = idmap_lookup_id(h, id);
+ if (he != 0) {
+ memcpy(name, he->ih_name, he->ih_namelen);
+ ret = he->ih_namelen;
+ up_read(&idmap->idmap_im_lock);
+ return ret;
+ }
+ up_read(&idmap->idmap_im_lock);
+
mutex_lock(&idmap->idmap_lock);
- mutex_lock(&idmap->idmap_im_lock);
+ /* Attempt lookup again in case we blocked
+ * because another attempt on this id
+ * was in progress.
+ */
+ down_read(&idmap->idmap_im_lock);
he = idmap_lookup_id(h, id);
if (he != 0) {
memcpy(name, he->ih_name, he->ih_namelen); ret = he->ih_namelen;
- goto out;
+ up_read(&idmap->idmap_im_lock);
+ mutex_unlock(&idmap->idmap_lock);
+ return ret;
}
memset(im, 0, sizeof(*im));
@@ -327,15 +357,15 @@ nfs_idmap_name(struct idmap *idmap, struct idmap_hashtable *h,
if (rpc_queue_upcall(idmap->idmap_dentry->d_inode, &msg) < 0) {
remove_wait_queue(&idmap->idmap_wq, &wq);
+ up_read(&idmap->idmap_im_lock);
goto out;
}
set_current_state(TASK_UNINTERRUPTIBLE);
- mutex_unlock(&idmap->idmap_im_lock);
+ up_read(&idmap->idmap_im_lock);
schedule();
current->state = TASK_RUNNING;
remove_wait_queue(&idmap->idmap_wq, &wq);
- mutex_lock(&idmap->idmap_im_lock);
if (im->im_status & IDMAP_STATUS_SUCCESS) {
if ((len = strnlen(im->im_name, IDMAP_NAMESZ)) == 0)
@@ -346,7 +376,6 @@ nfs_idmap_name(struct idmap *idmap, struct idmap_hashtable *h,
out:
memset(im, 0, sizeof(*im));
- mutex_unlock(&idmap->idmap_im_lock);
mutex_unlock(&idmap->idmap_lock);
return ret;
}
@@ -391,7 +420,7 @@ idmap_pipe_downcall(struct file *filp, const char __user *src, size_t mlen)
if (copy_from_user(&im_in, src, mlen) != 0)
return (-EFAULT);
- mutex_lock(&idmap->idmap_im_lock);
+ down_write(&idmap->idmap_im_lock);
ret = mlen;
im->im_status = im_in.im_status;
@@ -451,7 +480,7 @@ idmap_pipe_downcall(struct file *filp, const char __user *src, size_t mlen)
idmap_update_entry(he, im_in.im_name, namelen_in, im_in.im_id);
ret = mlen;
out:
- mutex_unlock(&idmap->idmap_im_lock);
+ up_write(&idmap->idmap_im_lock);
return ret;
}
@@ -463,10 +492,10 @@ idmap_pipe_destroy_msg(struct rpc_pipe_msg *msg)
if (msg->errno >= 0)
return;
- mutex_lock(&idmap->idmap_im_lock);
+ down_write(&idmap->idmap_im_lock);
im->im_status = IDMAP_STATUS_LOOKUPFAIL;
wake_up(&idmap->idmap_wq);
- mutex_unlock(&idmap->idmap_im_lock);
+ up_write(&idmap->idmap_im_lock);
}
/*
Signature · 18 lines
----------------------------------------------------------------- Company : Bull, Architect of an Open World TM (www.bull.com) Name : Aime Le Rouzic Mail : Bull - BP 208 - 38432 Echirolles Cedex - France E-Mail : [email protected] Phone : 33 (4) 76.29.75.51 Fax : 33 (4) 76.29.75.18 ----------------------------------------------------------------- ------------------------------------------------------------------------- This SF.net email is sponsored by DB2 Express Download DB2 Express C - the FREE version of DB2 express and take control of your XML. No limits. Just data. Click to get it now. http://sourceforge.net/powerbar/db2/ _______________________________________________ NFS maillist - [email protected] https://lists.sourceforge.net/lists/listinfo/nfs