From mboxrd@z Thu Jan 1 00:00:00 1970 From: Tao Ma Date: Fri, 21 Aug 2009 14:25:35 +0800 Subject: [Ocfs2-devel] [PATCH 17/41] ocfs2: Add CoW support. In-Reply-To: <20090821035547.GA20755@mail.oracle.com> References: <4A8A47DF.8020707@oracle.com> <1250576382-27080-17-git-send-email-tao.ma@oracle.com> <20090821005932.GE10558@mail.oracle.com> <4A8E00A2.1050902@oracle.com> <20090821025136.GK10558@mail.oracle.com> <20090821035547.GA20755@mail.oracle.com> Message-ID: <4A8E3DDF.8000300@oracle.com> List-Id: MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: ocfs2-devel@oss.oracle.com Joel Becker wrote: > On Thu, Aug 20, 2009 at 07:51:36PM -0700, Joel Becker wrote: >> I'm halfway through a modification of this code that splits out >> MAX_COW_BYTES from write_len. Let me finish it tomorrow. > > I just did it. What do you think? > > Joel > > diff --git a/fs/ocfs2/refcounttree.c b/fs/ocfs2/refcounttree.c > index d59860d..7790e1d 100644 > --- a/fs/ocfs2/refcounttree.c > +++ b/fs/ocfs2/refcounttree.c > @@ -2499,7 +2499,7 @@ out: > return ret; > } > > -#define MAX_COW_BYTES 1048576 > +#define MAX_CONTIG_BYTES 1048576 > /* > * Calculate out the start and number of virtual clusters we need to to CoW. > * > @@ -2508,9 +2508,8 @@ out: > * max_cpos is the place where we want to stop CoW intentionally. > * > * Normal we will start CoW from the beginning of extent record cotaining cpos. > - * And We will try to Cow as much clusters as we can until we reach > - * MAX_COW_BYTES. If the write_len is larger than MAX_COW_BYTES, we will > - * use that value as the maximum clusters. > + * We try to break up extents on boundaries of MAX_CONTIG_BYTES so that we > + * get good I/O from the resulting extent tree. > */ > static int ocfs2_refcount_cal_cow_clusters(struct inode *inode, > struct ocfs2_extent_list *el, > @@ -2525,10 +2524,11 @@ static int ocfs2_refcount_cal_cow_clusters(struct inode *inode, > struct buffer_head *eb_bh = NULL; > struct ocfs2_extent_block *eb = NULL; > struct ocfs2_extent_rec *rec; > - int max_clusters = ocfs2_clusters_for_bytes(inode->i_sb, MAX_COW_BYTES); > + int want_clusters; > + int contig_clusters = > + ocfs2_clusters_for_bytes(inode->i_sb, MAX_CONTIG_BYTES); > int leaf_clusters, rec_end = 0; > > - max_clusters = max_clusters < write_len ? write_len : max_clusters; > if (tree_height > 0) { > ret = ocfs2_find_leaf(INODE_CACHE(inode), el, cpos, &eb_bh); > if (ret) { > @@ -2587,53 +2587,84 @@ static int ocfs2_refcount_cal_cow_clusters(struct inode *inode, > leaf_clusters = rec_end - le32_to_cpu(rec->e_cpos); > } > > - if (*cow_len + leaf_clusters >= max_clusters) { > - if (*cow_len == 0) { > - /* > - * cpos is in a very large extent record. > - * So just split max_clusters from the > - * extent record. > - */ > - if ((rec_end - cpos) <= max_clusters) { > - /* > - * We can take max_clusters off > - * the end and cover all of our > - * write. > - */ > - *cow_start = rec_end - max_clusters; > - } else if ((*cow_start + max_clusters) > > - (cpos + write_len)) { > - /* > - * We can take max_clusters off > - * the front and cover all of > - * our write. > - */ > - /* NOOP, *cow_start is already set */ > - } else { > - /* > - * We're CoWing more data than > - * write_len for contiguousness, > - * but it doesn't fit at the > - * front or end of this extent. > - * Let's try to slice the extent > - * up nicely. Optimally, our > - * CoW region starts at a > - * multiple of max_clusters. If > - * that doesn't fit, we give up > - * and just CoW at cpos. > - */ > - *cow_start += > - (cpos - *cow_start) & > - ~(max_clusters - 1); > - if ((*cow_start + max_clusters) < > - (cpos + write_len)) > - *cow_start = cpos; > - } > - } > - *cow_len = max_clusters; > - break; > - } else > + /* > + * How many clusters do we actually need from > + * this extent? First we see how many we actually > + * need to complete the write. If that's smaller > + * than contig_clusters, we try for > + * contig_clustes. > + */ > + if (!*cow_len) > + want_clusters = write_len; > + else > + want_clusters = (cpos + write_len) - > + (*cow_start + *cow_len); > + if (want_clusters < contig_clusters) > + want_clusters = contig_clusters; > + > + /* > + * If the write does not cover the whole extent, we > + * need to calculate how we're going to split the extent. > + * We try to do it on contig_clusters boundaries. > + * > + * Any extent smaller than contig_clusters will be > + * CoWed in its entirety. > + */ > + if (leaf_clusters < contig_clusters) "<=" ? > *cow_len += leaf_clusters; > + else if (*cow_len || (*cow_start == cpos)) { > + /* > + * This extent needs to be CoW'd from its > + * beginning, so all we have to do is compute > + * how many clusters to grab. > + */ > + if (leaf_clusters < want_clusters) > + *cow_len += leaf_clusters; > + else > + *cow_len += want_clusters; > + } else if ((*cow_start + contig_clusters) > > + (cpos + write_len)) { ">="? otherwise, it looks good to me. btw, you will create a separate patch for this or I should integrate it into my original one? Regards, Tao > + /* > + * Breaking off contig_clusters at the front > + * of the extent will cover our write. That's > + * easy. > + */ > + *cow_len = contig_clusters; > + } else if ((rec_end - cpos) <= contig_clusters) { > + /* > + * Breaking off contig_clusters at the tail of > + * this extent will cover cpos. > + */ > + *cow_start = rec_end - cpos; > + *cow_len = contig_clusters; > + } else if ((rec_end - cpos) <= want_clusters) { > + /* > + * While we can't fit the entire write in this > + * extent, we know that the write goes from cpos > + * to the end of the extent. Break that off. > + */ > + *cow_start = cpos; > + *cow_len = rec_end - cpos; > + } else { > + /* > + * Ok, the entire write lives in the middle of > + * this extent. > + * Let's try to slice the extentup nicely. > + * Optimally, our CoW region starts at a > + * multiple of contig_clusters. If that doesn't > + * fit, we give up and just CoW at cpos. > + */ > + *cow_start += (cpos - *cow_start) & > + ~(contig_clusters - 1); > + if ((*cow_start + want_clusters) < > + (cpos + write_len)) > + *cow_start = cpos; > + *cow_len = want_clusters; > + } > + > + /* Have we covered our entire write yet? */ > + if ((*cow_start + *cow_len) >= (cpos + write_len)) > + break; > > /* > * If we reach the end of the extent block and don't get enough > >