xfs: add FITRIM support
authorChristoph Hellwig <hch@infradead.org>
Fri, 7 Jan 2011 13:02:04 +0000 (13:02 +0000)
committerAlex Elder <aelder@sgi.com>
Wed, 12 Jan 2011 02:28:29 +0000 (20:28 -0600)
Allow manual discards from userspace using the FITRIM ioctl.  This is not
intended to be run during normal workloads, as the freepsace btree walks
can cause large performance degradation.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Alex Elder <aelder@sgi.com>
fs/xfs/Makefile
fs/xfs/linux-2.6/xfs_discard.c [new file with mode: 0644]
fs/xfs/linux-2.6/xfs_discard.h [new file with mode: 0644]
fs/xfs/linux-2.6/xfs_ioctl.c
fs/xfs/linux-2.6/xfs_trace.h
fs/xfs/xfs_alloc.c
fs/xfs/xfs_alloc.h

index 0dce969d6cad61840430a40f3972de64a48bfdb0..faca449970995ab41a8fc9117efeceac63cf1991 100644 (file)
@@ -98,6 +98,7 @@ xfs-y                         += $(addprefix $(XFS_LINUX)/, \
                                   kmem.o \
                                   xfs_aops.o \
                                   xfs_buf.o \
+                                  xfs_discard.o \
                                   xfs_export.o \
                                   xfs_file.o \
                                   xfs_fs_subr.o \
diff --git a/fs/xfs/linux-2.6/xfs_discard.c b/fs/xfs/linux-2.6/xfs_discard.c
new file mode 100644 (file)
index 0000000..05201ae
--- /dev/null
@@ -0,0 +1,191 @@
+/*
+ * Copyright (C) 2010 Red Hat, Inc.
+ * All Rights Reserved.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License as
+ * published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope that it would be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write the Free Software Foundation,
+ * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
+ */
+#include "xfs.h"
+#include "xfs_sb.h"
+#include "xfs_inum.h"
+#include "xfs_log.h"
+#include "xfs_ag.h"
+#include "xfs_mount.h"
+#include "xfs_quota.h"
+#include "xfs_trans.h"
+#include "xfs_alloc_btree.h"
+#include "xfs_bmap_btree.h"
+#include "xfs_ialloc_btree.h"
+#include "xfs_btree.h"
+#include "xfs_inode.h"
+#include "xfs_alloc.h"
+#include "xfs_error.h"
+#include "xfs_discard.h"
+#include "xfs_trace.h"
+
+STATIC int
+xfs_trim_extents(
+       struct xfs_mount        *mp,
+       xfs_agnumber_t          agno,
+       xfs_fsblock_t           start,
+       xfs_fsblock_t           len,
+       xfs_fsblock_t           minlen,
+       __uint64_t              *blocks_trimmed)
+{
+       struct block_device     *bdev = mp->m_ddev_targp->bt_bdev;
+       struct xfs_btree_cur    *cur;
+       struct xfs_buf          *agbp;
+       struct xfs_perag        *pag;
+       int                     error;
+       int                     i;
+
+       pag = xfs_perag_get(mp, agno);
+
+       error = xfs_alloc_read_agf(mp, NULL, agno, 0, &agbp);
+       if (error || !agbp)
+               goto out_put_perag;
+
+       cur = xfs_allocbt_init_cursor(mp, NULL, agbp, agno, XFS_BTNUM_CNT);
+
+       /*
+        * Force out the log.  This means any transactions that might have freed
+        * space before we took the AGF buffer lock are now on disk, and the
+        * volatile disk cache is flushed.
+        */
+       xfs_log_force(mp, XFS_LOG_SYNC);
+
+       /*
+        * Look up the longest btree in the AGF and start with it.
+        */
+       error = xfs_alloc_lookup_le(cur, 0,
+                                   XFS_BUF_TO_AGF(agbp)->agf_longest, &i);
+       if (error)
+               goto out_del_cursor;
+
+       /*
+        * Loop until we are done with all extents that are large
+        * enough to be worth discarding.
+        */
+       while (i) {
+               xfs_agblock_t fbno;
+               xfs_extlen_t flen;
+
+               error = xfs_alloc_get_rec(cur, &fbno, &flen, &i);
+               if (error)
+                       goto out_del_cursor;
+               XFS_WANT_CORRUPTED_GOTO(i == 1, out_del_cursor);
+               ASSERT(flen <= XFS_BUF_TO_AGF(agbp)->agf_longest);
+
+               /*
+                * Too small?  Give up.
+                */
+               if (flen < minlen) {
+                       trace_xfs_discard_toosmall(mp, agno, fbno, flen);
+                       goto out_del_cursor;
+               }
+
+               /*
+                * If the extent is entirely outside of the range we are
+                * supposed to discard skip it.  Do not bother to trim
+                * down partially overlapping ranges for now.
+                */
+               if (XFS_AGB_TO_FSB(mp, agno, fbno) + flen < start ||
+                   XFS_AGB_TO_FSB(mp, agno, fbno) >= start + len) {
+                       trace_xfs_discard_exclude(mp, agno, fbno, flen);
+                       goto next_extent;
+               }
+
+               /*
+                * If any blocks in the range are still busy, skip the
+                * discard and try again the next time.
+                */
+               if (xfs_alloc_busy_search(mp, agno, fbno, flen)) {
+                       trace_xfs_discard_busy(mp, agno, fbno, flen);
+                       goto next_extent;
+               }
+
+               trace_xfs_discard_extent(mp, agno, fbno, flen);
+               error = -blkdev_issue_discard(bdev,
+                               XFS_AGB_TO_DADDR(mp, agno, fbno),
+                               XFS_FSB_TO_BB(mp, flen),
+                               GFP_NOFS, 0);
+               if (error)
+                       goto out_del_cursor;
+               *blocks_trimmed += flen;
+
+next_extent:
+               error = xfs_btree_decrement(cur, 0, &i);
+               if (error)
+                       goto out_del_cursor;
+       }
+
+out_del_cursor:
+       xfs_btree_del_cursor(cur, error ? XFS_BTREE_ERROR : XFS_BTREE_NOERROR);
+       xfs_buf_relse(agbp);
+out_put_perag:
+       xfs_perag_put(pag);
+       return error;
+}
+
+int
+xfs_ioc_trim(
+       struct xfs_mount                *mp,
+       struct fstrim_range __user      *urange)
+{
+       struct request_queue    *q = mp->m_ddev_targp->bt_bdev->bd_disk->queue;
+       unsigned int            granularity = q->limits.discard_granularity;
+       struct fstrim_range     range;
+       xfs_fsblock_t           start, len, minlen;
+       xfs_agnumber_t          start_agno, end_agno, agno;
+       __uint64_t              blocks_trimmed = 0;
+       int                     error, last_error = 0;
+
+       if (!capable(CAP_SYS_ADMIN))
+               return -XFS_ERROR(EPERM);
+       if (copy_from_user(&range, urange, sizeof(range)))
+               return -XFS_ERROR(EFAULT);
+
+       /*
+        * Truncating down the len isn't actually quite correct, but using
+        * XFS_B_TO_FSB would mean we trivially get overflows for values
+        * of ULLONG_MAX or slightly lower.  And ULLONG_MAX is the default
+        * used by the fstrim application.  In the end it really doesn't
+        * matter as trimming blocks is an advisory interface.
+        */
+       start = XFS_B_TO_FSBT(mp, range.start);
+       len = XFS_B_TO_FSBT(mp, range.len);
+       minlen = XFS_B_TO_FSB(mp, max_t(u64, granularity, range.minlen));
+
+       start_agno = XFS_FSB_TO_AGNO(mp, start);
+       if (start_agno >= mp->m_sb.sb_agcount)
+               return -XFS_ERROR(EINVAL);
+
+       end_agno = XFS_FSB_TO_AGNO(mp, start + len);
+       if (end_agno >= mp->m_sb.sb_agcount)
+               end_agno = mp->m_sb.sb_agcount - 1;
+
+       for (agno = start_agno; agno <= end_agno; agno++) {
+               error = -xfs_trim_extents(mp, agno, start, len, minlen,
+                                         &blocks_trimmed);
+               if (error)
+                       last_error = error;
+       }
+
+       if (last_error)
+               return last_error;
+
+       range.len = XFS_FSB_TO_B(mp, blocks_trimmed);
+       if (copy_to_user(urange, &range, sizeof(range)))
+               return -XFS_ERROR(EFAULT);
+       return 0;
+}
diff --git a/fs/xfs/linux-2.6/xfs_discard.h b/fs/xfs/linux-2.6/xfs_discard.h
new file mode 100644 (file)
index 0000000..e82b6dd
--- /dev/null
@@ -0,0 +1,8 @@
+#ifndef XFS_DISCARD_H
+#define XFS_DISCARD_H 1
+
+struct fstrim_range;
+
+extern int     xfs_ioc_trim(struct xfs_mount *, struct fstrim_range __user *);
+
+#endif /* XFS_DISCARD_H */
index ad442d9e392e480ab242f5d13922ee3ec0feab2e..b06ede1d0bed4562e60e4067e729b9209de65e1d 100644 (file)
@@ -39,6 +39,7 @@
 #include "xfs_dfrag.h"
 #include "xfs_fsops.h"
 #include "xfs_vnodeops.h"
+#include "xfs_discard.h"
 #include "xfs_quota.h"
 #include "xfs_inode_item.h"
 #include "xfs_export.h"
@@ -1294,6 +1295,8 @@ xfs_file_ioctl(
        trace_xfs_file_ioctl(ip);
 
        switch (cmd) {
+       case FITRIM:
+               return xfs_ioc_trim(mp, arg);
        case XFS_IOC_ALLOCSP:
        case XFS_IOC_FREESP:
        case XFS_IOC_RESVSP:
index 647af2a2e7aa8af2d14910e32beecfc00dbab515..2d0bcb479075fac542b176bbfb5a38ebc757a5dd 100644 (file)
@@ -1759,6 +1759,39 @@ DEFINE_LOG_RECOVER_INO_ITEM(xfs_log_recover_inode_recover);
 DEFINE_LOG_RECOVER_INO_ITEM(xfs_log_recover_inode_cancel);
 DEFINE_LOG_RECOVER_INO_ITEM(xfs_log_recover_inode_skip);
 
+DECLARE_EVENT_CLASS(xfs_discard_class,
+       TP_PROTO(struct xfs_mount *mp, xfs_agnumber_t agno,
+                xfs_agblock_t agbno, xfs_extlen_t len),
+       TP_ARGS(mp, agno, agbno, len),
+       TP_STRUCT__entry(
+               __field(dev_t, dev)
+               __field(xfs_agnumber_t, agno)
+               __field(xfs_agblock_t, agbno)
+               __field(xfs_extlen_t, len)
+       ),
+       TP_fast_assign(
+               __entry->dev = mp->m_super->s_dev;
+               __entry->agno = agno;
+               __entry->agbno = agbno;
+               __entry->len = len;
+       ),
+       TP_printk("dev %d:%d agno %u agbno %u len %u\n",
+                 MAJOR(__entry->dev), MINOR(__entry->dev),
+                 __entry->agno,
+                 __entry->agbno,
+                 __entry->len)
+)
+
+#define DEFINE_DISCARD_EVENT(name) \
+DEFINE_EVENT(xfs_discard_class, name, \
+       TP_PROTO(struct xfs_mount *mp, xfs_agnumber_t agno, \
+                xfs_agblock_t agbno, xfs_extlen_t len), \
+       TP_ARGS(mp, agno, agbno, len))
+DEFINE_DISCARD_EVENT(xfs_discard_extent);
+DEFINE_DISCARD_EVENT(xfs_discard_toosmall);
+DEFINE_DISCARD_EVENT(xfs_discard_exclude);
+DEFINE_DISCARD_EVENT(xfs_discard_busy);
+
 #endif /* _TRACE_XFS_H */
 
 #undef TRACE_INCLUDE_PATH
index fa8723f5870a113358b3bcd67214eba281b50b1f..f3227984a9bf815d554034ec4661bdcc37d6db36 100644 (file)
 #define        XFSA_FIXUP_BNO_OK       1
 #define        XFSA_FIXUP_CNT_OK       2
 
-static int
-xfs_alloc_busy_search(struct xfs_mount *mp, xfs_agnumber_t agno,
-                   xfs_agblock_t bno, xfs_extlen_t len);
-
 /*
  * Prototypes for per-ag allocation routines
  */
@@ -94,7 +90,7 @@ xfs_alloc_lookup_ge(
  * Lookup the first record less than or equal to [bno, len]
  * in the btree given by cur.
  */
-STATIC int                             /* error */
+int                                    /* error */
 xfs_alloc_lookup_le(
        struct xfs_btree_cur    *cur,   /* btree cursor */
        xfs_agblock_t           bno,    /* starting block of extent */
@@ -127,7 +123,7 @@ xfs_alloc_update(
 /*
  * Get the data from the pointed-to record.
  */
-STATIC int                             /* error */
+int                                    /* error */
 xfs_alloc_get_rec(
        struct xfs_btree_cur    *cur,   /* btree cursor */
        xfs_agblock_t           *bno,   /* output: starting block of extent */
@@ -2615,7 +2611,7 @@ restart:
  * will require a synchronous transaction, but it can still be
  * used to distinguish between a partial or exact match.
  */
-static int
+int
 xfs_alloc_busy_search(
        struct xfs_mount        *mp,
        xfs_agnumber_t          agno,
index 895009a97271fbbd27fcb63b97b9546f375a8778..0ab56b32c7eb70484ed4a7fdcfdbac3595c1ecab 100644 (file)
@@ -19,6 +19,7 @@
 #define        __XFS_ALLOC_H__
 
 struct xfs_buf;
+struct xfs_btree_cur;
 struct xfs_mount;
 struct xfs_perag;
 struct xfs_trans;
@@ -118,16 +119,16 @@ xfs_alloc_longest_free_extent(struct xfs_mount *mp,
                struct xfs_perag *pag);
 
 #ifdef __KERNEL__
-
 void
-xfs_alloc_busy_insert(xfs_trans_t *tp,
-               xfs_agnumber_t agno,
-               xfs_agblock_t bno,
-               xfs_extlen_t len);
+xfs_alloc_busy_insert(struct xfs_trans *tp, xfs_agnumber_t agno,
+       xfs_agblock_t bno, xfs_extlen_t len);
 
 void
 xfs_alloc_busy_clear(struct xfs_mount *mp, struct xfs_busy_extent *busyp);
 
+int
+xfs_alloc_busy_search(struct xfs_mount *mp, xfs_agnumber_t agno,
+       xfs_agblock_t bno, xfs_extlen_t len);
 #endif /* __KERNEL__ */
 
 /*
@@ -205,4 +206,18 @@ xfs_free_extent(
        xfs_fsblock_t   bno,    /* starting block number of extent */
        xfs_extlen_t    len);   /* length of extent */
 
+int                                    /* error */
+xfs_alloc_lookup_le(
+       struct xfs_btree_cur    *cur,   /* btree cursor */
+       xfs_agblock_t           bno,    /* starting block of extent */
+       xfs_extlen_t            len,    /* length of extent */
+       int                     *stat); /* success/failure */
+
+int                                    /* error */
+xfs_alloc_get_rec(
+       struct xfs_btree_cur    *cur,   /* btree cursor */
+       xfs_agblock_t           *bno,   /* output: starting block of extent */
+       xfs_extlen_t            *len,   /* output: length of extent */
+       int                     *stat); /* output: success/failure */
+
 #endif /* __XFS_ALLOC_H__ */