gfs2: read journal in large chunks
authorAbhi Das <adas@redhat.com>
Thu, 2 May 2019 19:17:40 +0000 (14:17 -0500)
committerAndreas Gruenbacher <agruenba@redhat.com>
Tue, 7 May 2019 21:39:15 +0000 (23:39 +0200)
Use bios to read in the journal into the address space of the journal inode
(jd_inode), sequentially and in large chunks.  This is faster for locating the
journal head that the previous binary search approach.  When performing
recovery, we keep the journal in the address space until recovery is done,
which further speeds up things.

Signed-off-by: Abhi Das <adas@redhat.com>
Signed-off-by: Andreas Gruenbacher <agruenba@redhat.com>
fs/gfs2/glops.c
fs/gfs2/log.c
fs/gfs2/lops.c
fs/gfs2/lops.h
fs/gfs2/ops_fstype.c
fs/gfs2/recovery.c
fs/gfs2/recovery.h
fs/gfs2/super.c

index 78510ab9183563e14c0c6de60761f87fdd234180..24ada3ccc525c8e7d6361edabf493fdd0b1c8c8e 100644 (file)
@@ -28,6 +28,7 @@
 #include "util.h"
 #include "trans.h"
 #include "dir.h"
+#include "lops.h"
 
 struct workqueue_struct *gfs2_freeze_wq;
 
@@ -531,7 +532,7 @@ static int freeze_go_xmote_bh(struct gfs2_glock *gl, struct gfs2_holder *gh)
        if (test_bit(SDF_JOURNAL_LIVE, &sdp->sd_flags)) {
                j_gl->gl_ops->go_inval(j_gl, DIO_METADATA);
 
-               error = gfs2_find_jhead(sdp->sd_jdesc, &head);
+               error = gfs2_find_jhead(sdp->sd_jdesc, &head, false);
                if (error)
                        gfs2_consist(sdp);
                if (!(head.lh_flags & GFS2_LOG_HEAD_UNMOUNT))
index a7febb4bd400465942249ec284867324cd2b73e6..a2e1df488df04a305d2a8eca84327a2461e158bf 100644 (file)
@@ -744,7 +744,7 @@ void gfs2_write_log_header(struct gfs2_sbd *sdp, struct gfs2_jdesc *jd,
        lh->lh_crc = cpu_to_be32(crc);
 
        gfs2_log_write(sdp, page, sb->s_blocksize, 0, dblock);
-       gfs2_log_submit_bio(&sdp->sd_log_bio, REQ_OP_WRITE, op_flags);
+       gfs2_log_submit_bio(&sdp->sd_log_bio, REQ_OP_WRITE | op_flags);
        log_flush_wait(sdp);
 }
 
@@ -821,7 +821,7 @@ void gfs2_log_flush(struct gfs2_sbd *sdp, struct gfs2_glock *gl, u32 flags)
 
        gfs2_ordered_write(sdp);
        lops_before_commit(sdp, tr);
-       gfs2_log_submit_bio(&sdp->sd_log_bio, REQ_OP_WRITE, 0);
+       gfs2_log_submit_bio(&sdp->sd_log_bio, REQ_OP_WRITE);
 
        if (sdp->sd_log_head != sdp->sd_log_flush_head) {
                log_flush_wait(sdp);
index 6af6a3cea9679ab769c18d5a573b379fc605baaf..ce048a9e058d7cd58784514f22625222c5058a03 100644 (file)
@@ -17,7 +17,9 @@
 #include <linux/bio.h>
 #include <linux/fs.h>
 #include <linux/list_sort.h>
+#include <linux/blkdev.h>
 
+#include "bmap.h"
 #include "dir.h"
 #include "gfs2.h"
 #include "incore.h"
@@ -194,7 +196,6 @@ static void gfs2_end_log_write_bh(struct gfs2_sbd *sdp,
 /**
  * gfs2_end_log_write - end of i/o to the log
  * @bio: The bio
- * @error: Status of i/o request
  *
  * Each bio_vec contains either data from the pagecache or data
  * relating to the log itself. Here we iterate over the bio_vec
@@ -232,20 +233,19 @@ static void gfs2_end_log_write(struct bio *bio)
 /**
  * gfs2_log_submit_bio - Submit any pending log bio
  * @biop: Address of the bio pointer
- * @op: REQ_OP
- * @op_flags: req_flag_bits
+ * @opf: REQ_OP | op_flags
  *
  * Submit any pending part-built or full bio to the block device. If
  * there is no pending bio, then this is a no-op.
  */
 
-void gfs2_log_submit_bio(struct bio **biop, int op, int op_flags)
+void gfs2_log_submit_bio(struct bio **biop, int opf)
 {
        struct bio *bio = *biop;
        if (bio) {
                struct gfs2_sbd *sdp = bio->bi_private;
                atomic_inc(&sdp->sd_log_in_flight);
-               bio_set_op_attrs(bio, op, op_flags);
+               bio->bi_opf = opf;
                submit_bio(bio);
                *biop = NULL;
        }
@@ -306,7 +306,7 @@ static struct bio *gfs2_log_get_bio(struct gfs2_sbd *sdp, u64 blkno,
                nblk >>= sdp->sd_fsb2bb_shift;
                if (blkno == nblk && !flush)
                        return bio;
-               gfs2_log_submit_bio(biop, op, 0);
+               gfs2_log_submit_bio(biop, op);
        }
 
        *biop = gfs2_log_alloc_bio(sdp, blkno, end_io);
@@ -377,6 +377,206 @@ void gfs2_log_write_page(struct gfs2_sbd *sdp, struct page *page)
                       gfs2_log_bmap(sdp));
 }
 
+/**
+ * gfs2_end_log_read - end I/O callback for reads from the log
+ * @bio: The bio
+ *
+ * Simply unlock the pages in the bio. The main thread will wait on them and
+ * process them in order as necessary.
+ */
+
+static void gfs2_end_log_read(struct bio *bio)
+{
+       struct page *page;
+       struct bio_vec *bvec;
+       int i;
+       struct bvec_iter_all iter_all;
+
+       bio_for_each_segment_all(bvec, bio, i, iter_all) {
+               page = bvec->bv_page;
+               if (bio->bi_status) {
+                       int err = blk_status_to_errno(bio->bi_status);
+
+                       SetPageError(page);
+                       mapping_set_error(page->mapping, err);
+               }
+               unlock_page(page);
+       }
+
+       bio_put(bio);
+}
+
+/**
+ * gfs2_jhead_pg_srch - Look for the journal head in a given page.
+ * @jd: The journal descriptor
+ * @page: The page to look in
+ *
+ * Returns: 1 if found, 0 otherwise.
+ */
+
+static bool gfs2_jhead_pg_srch(struct gfs2_jdesc *jd,
+                             struct gfs2_log_header_host *head,
+                             struct page *page)
+{
+       struct gfs2_sbd *sdp = GFS2_SB(jd->jd_inode);
+       struct gfs2_log_header_host uninitialized_var(lh);
+       void *kaddr = kmap_atomic(page);
+       unsigned int offset;
+       bool ret = false;
+
+       for (offset = 0; offset < PAGE_SIZE; offset += sdp->sd_sb.sb_bsize) {
+               if (!__get_log_header(sdp, kaddr + offset, 0, &lh)) {
+                       if (lh.lh_sequence > head->lh_sequence)
+                               *head = lh;
+                       else {
+                               ret = true;
+                               break;
+                       }
+               }
+       }
+       kunmap_atomic(kaddr);
+       return ret;
+}
+
+/**
+ * gfs2_jhead_process_page - Search/cleanup a page
+ * @jd: The journal descriptor
+ * @index: Index of the page to look into
+ * @done: If set, perform only cleanup, else search and set if found.
+ *
+ * Find the page with 'index' in the journal's mapping. Search the page for
+ * the journal head if requested (cleanup == false). Release refs on the
+ * page so the page cache can reclaim it (put_page() twice). We grabbed a
+ * reference on this page two times, first when we did a find_or_create_page()
+ * to obtain the page to add it to the bio and second when we do a
+ * find_get_page() here to get the page to wait on while I/O on it is being
+ * completed.
+ * This function is also used to free up a page we might've grabbed but not
+ * used. Maybe we added it to a bio, but not submitted it for I/O. Or we
+ * submitted the I/O, but we already found the jhead so we only need to drop
+ * our references to the page.
+ */
+
+static void gfs2_jhead_process_page(struct gfs2_jdesc *jd, unsigned long index,
+                                   struct gfs2_log_header_host *head,
+                                   bool *done)
+{
+       struct page *page;
+
+       page = find_get_page(jd->jd_inode->i_mapping, index);
+       wait_on_page_locked(page);
+
+       if (PageError(page))
+               *done = true;
+
+       if (!*done)
+               *done = gfs2_jhead_pg_srch(jd, head, page);
+
+       put_page(page); /* Once for find_get_page */
+       put_page(page); /* Once more for find_or_create_page */
+}
+
+/**
+ * gfs2_find_jhead - find the head of a log
+ * @jd: The journal descriptor
+ * @head: The log descriptor for the head of the log is returned here
+ *
+ * Do a search of a journal by reading it in large chunks using bios and find
+ * the valid log entry with the highest sequence number.  (i.e. the log head)
+ *
+ * Returns: 0 on success, errno otherwise
+ */
+int gfs2_find_jhead(struct gfs2_jdesc *jd, struct gfs2_log_header_host *head,
+                   bool keep_cache)
+{
+       struct gfs2_sbd *sdp = GFS2_SB(jd->jd_inode);
+       struct address_space *mapping = jd->jd_inode->i_mapping;
+       unsigned int block = 0, blocks_submitted = 0, blocks_read = 0;
+       unsigned int bsize = sdp->sd_sb.sb_bsize;
+       unsigned int bsize_shift = sdp->sd_sb.sb_bsize_shift;
+       unsigned int shift = PAGE_SHIFT - bsize_shift;
+       unsigned int readhead_blocks = BIO_MAX_PAGES << shift;
+       struct gfs2_journal_extent *je;
+       int sz, ret = 0;
+       struct bio *bio = NULL;
+       struct page *page = NULL;
+       bool done = false;
+       errseq_t since;
+
+       memset(head, 0, sizeof(*head));
+       if (list_empty(&jd->extent_list))
+               gfs2_map_journal_extents(sdp, jd);
+
+       since = filemap_sample_wb_err(mapping);
+       list_for_each_entry(je, &jd->extent_list, list) {
+               for (; block < je->lblock + je->blocks; block++) {
+                       u64 dblock;
+
+                       if (!page) {
+                               page = find_or_create_page(mapping,
+                                               block >> shift, GFP_NOFS);
+                               if (!page) {
+                                       ret = -ENOMEM;
+                                       done = true;
+                                       goto out;
+                               }
+                       }
+
+                       if (bio) {
+                               unsigned int off;
+
+                               off = (block << bsize_shift) & ~PAGE_MASK;
+                               sz = bio_add_page(bio, page, bsize, off);
+                               if (sz == bsize) { /* block added */
+                                       if (off + bsize == PAGE_SIZE) {
+                                               page = NULL;
+                                               goto page_added;
+                                       }
+                                       continue;
+                               }
+                               blocks_submitted = block + 1;
+                               submit_bio(bio);
+                               bio = NULL;
+                       }
+
+                       dblock = je->dblock + (block - je->lblock);
+                       bio = gfs2_log_alloc_bio(sdp, dblock, gfs2_end_log_read);
+                       bio->bi_opf = REQ_OP_READ;
+                       sz = bio_add_page(bio, page, bsize, 0);
+                       gfs2_assert_warn(sdp, sz == bsize);
+                       if (bsize == PAGE_SIZE)
+                               page = NULL;
+
+page_added:
+                       if (blocks_submitted < blocks_read + readhead_blocks) {
+                               /* Keep at least one bio in flight */
+                               continue;
+                       }
+
+                       gfs2_jhead_process_page(jd, blocks_read >> shift, head, &done);
+                       blocks_read += PAGE_SIZE >> bsize_shift;
+                       if (done)
+                               goto out;  /* found */
+               }
+       }
+
+out:
+       if (bio)
+               submit_bio(bio);
+       while (blocks_read < block) {
+               gfs2_jhead_process_page(jd, blocks_read >> shift, head, &done);
+               blocks_read += PAGE_SIZE >> bsize_shift;
+       }
+
+       if (!ret)
+               ret = filemap_check_wb_err(mapping, since);
+
+       if (!keep_cache)
+               truncate_inode_pages(mapping, 0);
+
+       return ret;
+}
+
 static struct page *gfs2_get_log_desc(struct gfs2_sbd *sdp, u32 ld_type,
                                      u32 ld_length, u32 ld_data1)
 {
index 320fbf28d2fba01610fef0a3f1f100250b0ed48a..f195ffb435ac6fa83bc568bc57b78639403eaf28 100644 (file)
@@ -25,8 +25,10 @@ extern u64 gfs2_log_bmap(struct gfs2_sbd *sdp);
 extern void gfs2_log_write(struct gfs2_sbd *sdp, struct page *page,
                           unsigned size, unsigned offset, u64 blkno);
 extern void gfs2_log_write_page(struct gfs2_sbd *sdp, struct page *page);
-extern void gfs2_log_submit_bio(struct bio **biop, int op, int op_flags);
+extern void gfs2_log_submit_bio(struct bio **biop, int opf);
 extern void gfs2_pin(struct gfs2_sbd *sdp, struct buffer_head *bh);
+extern int gfs2_find_jhead(struct gfs2_jdesc *jd,
+                          struct gfs2_log_header_host *head, bool keep_cache);
 
 static inline unsigned int buf_limit(struct gfs2_sbd *sdp)
 {
index abfaecde0e3dc8007768a5580674163d627e2d93..46f6615eaf124972eb134be82f24dfe0df0203be 100644 (file)
@@ -41,6 +41,7 @@
 #include "dir.h"
 #include "meta_io.h"
 #include "trace_gfs2.h"
+#include "lops.h"
 
 #define DO 0
 #define UNDO 1
@@ -616,7 +617,7 @@ static int check_journal_clean(struct gfs2_sbd *sdp, struct gfs2_jdesc *jd)
                fs_err(sdp, "Error checking journal for spectator mount.\n");
                goto out_unlock;
        }
-       error = gfs2_find_jhead(jd, &head);
+       error = gfs2_find_jhead(jd, &head, false);
        if (error) {
                fs_err(sdp, "Error parsing journal for spectator mount.\n");
                goto out_unlock;
index fa575d1676b9b9144bed3f8561b088d5540ffc40..389b3ef77e20fea401fdcb3d1bbe69b83d7ffa4c 100644 (file)
@@ -181,129 +181,6 @@ static int get_log_header(struct gfs2_jdesc *jd, unsigned int blk,
        return error;
 }
 
-/**
- * find_good_lh - find a good log header
- * @jd: the journal
- * @blk: the segment to start searching from
- * @lh: the log header to fill in
- * @forward: if true search forward in the log, else search backward
- *
- * Call get_log_header() to get a log header for a segment, but if the
- * segment is bad, either scan forward or backward until we find a good one.
- *
- * Returns: errno
- */
-
-static int find_good_lh(struct gfs2_jdesc *jd, unsigned int *blk,
-                       struct gfs2_log_header_host *head)
-{
-       unsigned int orig_blk = *blk;
-       int error;
-
-       for (;;) {
-               error = get_log_header(jd, *blk, head);
-               if (error <= 0)
-                       return error;
-
-               if (++*blk == jd->jd_blocks)
-                       *blk = 0;
-
-               if (*blk == orig_blk) {
-                       gfs2_consist_inode(GFS2_I(jd->jd_inode));
-                       return -EIO;
-               }
-       }
-}
-
-/**
- * jhead_scan - make sure we've found the head of the log
- * @jd: the journal
- * @head: this is filled in with the log descriptor of the head
- *
- * At this point, seg and lh should be either the head of the log or just
- * before.  Scan forward until we find the head.
- *
- * Returns: errno
- */
-
-static int jhead_scan(struct gfs2_jdesc *jd, struct gfs2_log_header_host *head)
-{
-       unsigned int blk = head->lh_blkno;
-       struct gfs2_log_header_host lh;
-       int error;
-
-       for (;;) {
-               if (++blk == jd->jd_blocks)
-                       blk = 0;
-
-               error = get_log_header(jd, blk, &lh);
-               if (error < 0)
-                       return error;
-               if (error == 1)
-                       continue;
-
-               if (lh.lh_sequence == head->lh_sequence) {
-                       gfs2_consist_inode(GFS2_I(jd->jd_inode));
-                       return -EIO;
-               }
-               if (lh.lh_sequence < head->lh_sequence)
-                       break;
-
-               *head = lh;
-       }
-
-       return 0;
-}
-
-/**
- * gfs2_find_jhead - find the head of a log
- * @jd: the journal
- * @head: the log descriptor for the head of the log is returned here
- *
- * Do a binary search of a journal and find the valid log entry with the
- * highest sequence number.  (i.e. the log head)
- *
- * Returns: errno
- */
-
-int gfs2_find_jhead(struct gfs2_jdesc *jd, struct gfs2_log_header_host *head)
-{
-       struct gfs2_log_header_host lh_1, lh_m;
-       u32 blk_1, blk_2, blk_m;
-       int error;
-
-       blk_1 = 0;
-       blk_2 = jd->jd_blocks - 1;
-
-       for (;;) {
-               blk_m = (blk_1 + blk_2) / 2;
-
-               error = find_good_lh(jd, &blk_1, &lh_1);
-               if (error)
-                       return error;
-
-               error = find_good_lh(jd, &blk_m, &lh_m);
-               if (error)
-                       return error;
-
-               if (blk_1 == blk_m || blk_m == blk_2)
-                       break;
-
-               if (lh_1.lh_sequence <= lh_m.lh_sequence)
-                       blk_1 = blk_m;
-               else
-                       blk_2 = blk_m;
-       }
-
-       error = jhead_scan(jd, &lh_1);
-       if (error)
-               return error;
-
-       *head = lh_1;
-
-       return error;
-}
-
 /**
  * foreach_descriptor - go through the active part of the log
  * @jd: the journal
@@ -469,7 +346,7 @@ void gfs2_recover_func(struct work_struct *work)
        if (error)
                goto fail_gunlock_ji;
 
-       error = gfs2_find_jhead(jd, &head);
+       error = gfs2_find_jhead(jd, &head, true);
        if (error)
                goto fail_gunlock_ji;
        t_jhd = ktime_get();
index 5932d4b6f43e6e13cee6904173083a490df57dec..1831a1974c8cb26a575502cde7087b9388669b02 100644 (file)
@@ -27,8 +27,6 @@ extern int gfs2_revoke_add(struct gfs2_jdesc *jd, u64 blkno, unsigned int where)
 extern int gfs2_revoke_check(struct gfs2_jdesc *jd, u64 blkno, unsigned int where);
 extern void gfs2_revoke_clean(struct gfs2_jdesc *jd);
 
-extern int gfs2_find_jhead(struct gfs2_jdesc *jd,
-                   struct gfs2_log_header_host *head);
 extern int gfs2_recover_journal(struct gfs2_jdesc *gfs2_jd, bool wait);
 extern void gfs2_recover_func(struct work_struct *work);
 extern int __get_log_header(struct gfs2_sbd *sdp,
index 5b148cf032935218847e964c1cc757775c16d41d..fbf6b1fd330b315f761eabd58ab66b87c5bae1bc 100644 (file)
@@ -45,6 +45,7 @@
 #include "util.h"
 #include "sys.h"
 #include "xattr.h"
+#include "lops.h"
 
 #define args_neq(a1, a2, x) ((a1)->ar_##x != (a2)->ar_##x)
 
@@ -425,7 +426,7 @@ int gfs2_make_fs_rw(struct gfs2_sbd *sdp)
 
        j_gl->gl_ops->go_inval(j_gl, DIO_METADATA);
 
-       error = gfs2_find_jhead(sdp->sd_jdesc, &head);
+       error = gfs2_find_jhead(sdp->sd_jdesc, &head, false);
        if (error)
                goto fail;
 
@@ -680,7 +681,7 @@ static int gfs2_lock_fs_check_clean(struct gfs2_sbd *sdp,
                error = gfs2_jdesc_check(jd);
                if (error)
                        break;
-               error = gfs2_find_jhead(jd, &lh);
+               error = gfs2_find_jhead(jd, &lh, false);
                if (error)
                        break;
                if (!(lh.lh_flags & GFS2_LOG_HEAD_UNMOUNT)) {