pNFS/flexfiles: Improve merging of errors in LAYOUTRETURN
authorTrond Myklebust <trond.myklebust@primarydata.com>
Thu, 21 Jan 2016 19:48:49 +0000 (14:48 -0500)
committerTrond Myklebust <trond.myklebust@primarydata.com>
Thu, 21 Jan 2016 20:49:40 +0000 (15:49 -0500)
When we hit 22 errors, we start to overflow the memory buffers allocated
to the LAYOUTRETURN errors. The issue is that currently, RPC call reply
ordering determines how successful we are in merging errors that refer
to contiguous READ or WRITE requests.

Fix is to use an insertion sort to help detect contiguity.

Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
fs/nfs/flexfilelayout/flexfilelayoutdev.c

index e125e55de86daebcbba9f2f907aca05fa27a0a1a..bcf79aad86fc3109b7a976704241e198c061b738 100644 (file)
@@ -218,63 +218,55 @@ static void extend_ds_error(struct nfs4_ff_layout_ds_err *err,
        err->length = end - err->offset;
 }
 
-static bool ds_error_can_merge(struct nfs4_ff_layout_ds_err *err,  u64 offset,
-                              u64 length, int status, enum nfs_opnum4 opnum,
-                              nfs4_stateid *stateid,
-                              struct nfs4_deviceid *deviceid)
+static int
+ff_ds_error_match(const struct nfs4_ff_layout_ds_err *e1,
+               const struct nfs4_ff_layout_ds_err *e2)
 {
-       return err->status == status && err->opnum == opnum &&
-              nfs4_stateid_match(&err->stateid, stateid) &&
-              !memcmp(&err->deviceid, deviceid, sizeof(*deviceid)) &&
-              end_offset(err->offset, err->length) >= offset &&
-              err->offset <= end_offset(offset, length);
-}
-
-static bool merge_ds_error(struct nfs4_ff_layout_ds_err *old,
-                          struct nfs4_ff_layout_ds_err *new)
-{
-       if (!ds_error_can_merge(old, new->offset, new->length, new->status,
-                               new->opnum, &new->stateid, &new->deviceid))
-               return false;
-
-       extend_ds_error(old, new->offset, new->length);
-       return true;
+       int ret;
+
+       if (e1->opnum != e2->opnum)
+               return e1->opnum < e2->opnum ? -1 : 1;
+       if (e1->status != e2->status)
+               return e1->status < e2->status ? -1 : 1;
+       ret = memcmp(&e1->stateid, &e2->stateid, sizeof(e1->stateid));
+       if (ret != 0)
+               return ret;
+       ret = memcmp(&e1->deviceid, &e2->deviceid, sizeof(e1->deviceid));
+       if (ret != 0)
+               return ret;
+       if (end_offset(e1->offset, e1->length) < e2->offset)
+               return -1;
+       if (e1->offset > end_offset(e2->offset, e2->length))
+               return 1;
+       /* If ranges overlap or are contiguous, they are the same */
+       return 0;
 }
 
-static bool
+static void
 ff_layout_add_ds_error_locked(struct nfs4_flexfile_layout *flo,
                              struct nfs4_ff_layout_ds_err *dserr)
 {
-       struct nfs4_ff_layout_ds_err *err;
-
-       list_for_each_entry(err, &flo->error_list, list) {
-               if (merge_ds_error(err, dserr)) {
-                       return true;
-               }
-       }
-
-       list_add(&dserr->list, &flo->error_list);
-       return false;
-}
-
-static bool
-ff_layout_update_ds_error(struct nfs4_flexfile_layout *flo, u64 offset,
-                         u64 length, int status, enum nfs_opnum4 opnum,
-                         nfs4_stateid *stateid, struct nfs4_deviceid *deviceid)
-{
-       bool found = false;
-       struct nfs4_ff_layout_ds_err *err;
-
-       list_for_each_entry(err, &flo->error_list, list) {
-               if (ds_error_can_merge(err, offset, length, status, opnum,
-                                      stateid, deviceid)) {
-                       found = true;
-                       extend_ds_error(err, offset, length);
+       struct nfs4_ff_layout_ds_err *err, *tmp;
+       struct list_head *head = &flo->error_list;
+       int match;
+
+       /* Do insertion sort w/ merges */
+       list_for_each_entry_safe(err, tmp, &flo->error_list, list) {
+               match = ff_ds_error_match(err, dserr);
+               if (match < 0)
+                       continue;
+               if (match > 0) {
+                       /* Add entry "dserr" _before_ entry "err" */
+                       head = &err->list;
                        break;
                }
+               /* Entries match, so merge "err" into "dserr" */
+               extend_ds_error(dserr, err->offset, err->length);
+               list_del(&err->list);
+               kfree(err);
        }
 
-       return found;
+       list_add_tail(&dserr->list, head);
 }
 
 int ff_layout_track_ds_error(struct nfs4_flexfile_layout *flo,
@@ -283,7 +275,6 @@ int ff_layout_track_ds_error(struct nfs4_flexfile_layout *flo,
                             gfp_t gfp_flags)
 {
        struct nfs4_ff_layout_ds_err *dserr;
-       bool needfree;
 
        if (status == 0)
                return 0;
@@ -291,14 +282,6 @@ int ff_layout_track_ds_error(struct nfs4_flexfile_layout *flo,
        if (mirror->mirror_ds == NULL)
                return -EINVAL;
 
-       spin_lock(&flo->generic_hdr.plh_inode->i_lock);
-       if (ff_layout_update_ds_error(flo, offset, length, status, opnum,
-                                     &mirror->stateid,
-                                     &mirror->mirror_ds->id_node.deviceid)) {
-               spin_unlock(&flo->generic_hdr.plh_inode->i_lock);
-               return 0;
-       }
-       spin_unlock(&flo->generic_hdr.plh_inode->i_lock);
        dserr = kmalloc(sizeof(*dserr), gfp_flags);
        if (!dserr)
                return -ENOMEM;
@@ -313,10 +296,8 @@ int ff_layout_track_ds_error(struct nfs4_flexfile_layout *flo,
               NFS4_DEVICEID4_SIZE);
 
        spin_lock(&flo->generic_hdr.plh_inode->i_lock);
-       needfree = ff_layout_add_ds_error_locked(flo, dserr);
+       ff_layout_add_ds_error_locked(flo, dserr);
        spin_unlock(&flo->generic_hdr.plh_inode->i_lock);
-       if (needfree)
-               kfree(dserr);
 
        return 0;
 }