xfs: add online discard support

Now that we have reliably tracking of deleted extents in a transaction we can easily implement "online" discard support which calls blkdev_issue_discard once a transaction commits. The actual discard is a two stage operation as we first have to mark the busy extent as not available for reuse before we can start the actual discard. Note that we don't bother supporting discard for the non-delaylog mode. Signed-off-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Alex Elder <aelder@sgi.com>

xfs: add online discard support
Now that we have reliably tracking of deleted extents in a transaction we can easily implement "online" discard support which calls blkdev_issue_discard once a transaction commits. The actual discard is a two stage operation as we first have to mark the busy extent as not available for reuse before we can start the actual discard. Note that we don't bother supporting discard for the non-delaylog mode. Signed-off-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Alex Elder <aelder@sgi.com>
Christoph Hellwig · Alex Elder
1 parent bf59170a66
Showing 10 changed files with 96 additions and 8 deletions Side-by-side Diff
Documentation/filesystems/xfs.txt
fs/xfs/linux-2.6/xfs_discard.c
fs/xfs/linux-2.6/xfs_discard.h
fs/xfs/linux-2.6/xfs_super.c
fs/xfs/xfs_ag.h
fs/xfs/xfs_alloc.c
fs/xfs/xfs_alloc.h
fs/xfs/xfs_log_cil.c
fs/xfs/xfs_mount.h
fs/xfs/xfs_trans.c
@@ -39,6 +39,12 @@
 	drive level write caching to be enabled, for devices that
 	support write barriers.
  
+  discard
+	Issue command to let the block device reclaim space freed by the
+	filesystem.  This is useful for SSD devices, thinly provisioned
+	LUNs and virtual machine images, but may have a performance
+	impact.  This option is incompatible with the nodelaylog option.
+
   dmapi
 	Enable the DMAPI (Data Management API) event callouts.
 	Use with the "mtpt" option.
@@ -191,4 +191,33 @@
 		return -XFS_ERROR(EFAULT);
 	return 0;
 }
+
+int
+xfs_discard_extents(
+	struct xfs_mount	*mp,
+	struct list_head	*list)
+{
+	struct xfs_busy_extent	*busyp;
+	int			error = 0;
+
+	list_for_each_entry(busyp, list, list) {
+		trace_xfs_discard_extent(mp, busyp->agno, busyp->bno,
+					 busyp->length);
+
+		error = -blkdev_issue_discard(mp->m_ddev_targp->bt_bdev,
+				XFS_AGB_TO_DADDR(mp, busyp->agno, busyp->bno),
+				XFS_FSB_TO_BB(mp, busyp->length),
+				GFP_NOFS, 0);
+		if (error && error != EOPNOTSUPP) {
+			xfs_info(mp,
+	 "discard failed for extent [0x%llu,%u], error %d",
+				 (unsigned long long)busyp->bno,
+				 busyp->length,
+				 error);
+			return error;
+		}
+	}
+
+	return 0;
+}
@@ -2,8 +2,10 @@
 #define XFS_DISCARD_H 1
  
 struct fstrim_range;
+struct list_head;
  
 extern int	xfs_ioc_trim(struct xfs_mount *, struct fstrim_range __user *);
+extern int	xfs_discard_extents(struct xfs_mount *, struct list_head *);
  
 #endif /* XFS_DISCARD_H */
@@ -110,8 +110,10 @@
 #define MNTOPT_GQUOTANOENF "gqnoenforce"/* group quota limit enforcement */
 #define MNTOPT_PQUOTANOENF "pqnoenforce"/* project quota limit enforcement */
 #define MNTOPT_QUOTANOENF  "qnoenforce"	/* same as uqnoenforce */
-#define MNTOPT_DELAYLOG   "delaylog"	/* Delayed loging enabled */
-#define MNTOPT_NODELAYLOG "nodelaylog"	/* Delayed loging disabled */
+#define MNTOPT_DELAYLOG    "delaylog"	/* Delayed logging enabled */
+#define MNTOPT_NODELAYLOG  "nodelaylog"	/* Delayed logging disabled */
+#define MNTOPT_DISCARD	   "discard"	/* Discard unused blocks */
+#define MNTOPT_NODISCARD   "nodiscard"	/* Do not discard unused blocks */
  
 /*
  * Table driven mount option parser.
@@ -355,6 +357,10 @@
 			mp->m_flags |= XFS_MOUNT_DELAYLOG;
 		} else if (!strcmp(this_char, MNTOPT_NODELAYLOG)) {
 			mp->m_flags &= ~XFS_MOUNT_DELAYLOG;
+		} else if (!strcmp(this_char, MNTOPT_DISCARD)) {
+			mp->m_flags |= XFS_MOUNT_DISCARD;
+		} else if (!strcmp(this_char, MNTOPT_NODISCARD)) {
+			mp->m_flags &= ~XFS_MOUNT_DISCARD;
 		} else if (!strcmp(this_char, "ihashsize")) {
 			xfs_warn(mp,
 	"ihashsize no longer used, option is deprecated.");
@@ -388,6 +394,13 @@
 		return EINVAL;
 	}
  
+	if ((mp->m_flags & XFS_MOUNT_DISCARD) &&
+	    !(mp->m_flags & XFS_MOUNT_DELAYLOG)) {
+		xfs_warn(mp,
+	"the discard option is incompatible with the nodelaylog option");
+		return EINVAL;
+	}
+
 #ifndef CONFIG_XFS_QUOTA
 	if (XFS_IS_QUOTA_RUNNING(mp)) {
 		xfs_warn(mp, "quota support not available in this kernel.");
@@ -488,6 +501,7 @@
 		{ XFS_MOUNT_FILESTREAMS,	"," MNTOPT_FILESTREAM },
 		{ XFS_MOUNT_GRPID,		"," MNTOPT_GRPID },
 		{ XFS_MOUNT_DELAYLOG,		"," MNTOPT_DELAYLOG },
+		{ XFS_MOUNT_DISCARD,		"," MNTOPT_DISCARD },
 		{ 0, NULL }
 	};
 	static struct proc_xfs_info xfs_info_unset[] = {
@@ -187,6 +187,8 @@
 	xfs_agnumber_t	agno;
 	xfs_agblock_t	bno;
 	xfs_extlen_t	length;
+	unsigned int	flags;
+#define XFS_ALLOC_BUSY_DISCARDED	0x01	/* undergoing a discard op. */
 };
  
 /*
@@ -2609,6 +2609,18 @@
 	xfs_agblock_t		bend = bbno + busyp->length;
  
 	/*
+	 * This extent is currently being discarded.  Give the thread
+	 * performing the discard a chance to mark the extent unbusy
+	 * and retry.
+	 */
+	if (busyp->flags & XFS_ALLOC_BUSY_DISCARDED) {
+		spin_unlock(&pag->pagb_lock);
+		delay(1);
+		spin_lock(&pag->pagb_lock);
+		return false;
+	}
+
+	/*
 	 * If there is a busy extent overlapping a user allocation, we have
 	 * no choice but to force the log and retry the search.
 	 *
@@ -2813,7 +2825,8 @@
 		 * If this is a metadata allocation, try to reuse the busy
 		 * extent instead of trimming the allocation.
 		 */
-		if (!args->userdata) {
+		if (!args->userdata &&
+		    !(busyp->flags & XFS_ALLOC_BUSY_DISCARDED)) {
 			if (!xfs_alloc_busy_update_extent(args->mp, args->pag,
 							  busyp, fbno, flen,
 							  false))
  
@@ -2979,10 +2992,16 @@
 	kmem_free(busyp);
 }
  
+/*
+ * Remove all extents on the passed in list from the busy extents tree.
+ * If do_discard is set skip extents that need to be discarded, and mark
+ * these as undergoing a discard operation instead.
+ */
 void
 xfs_alloc_busy_clear(
 	struct xfs_mount	*mp,
-	struct list_head	*list)
+	struct list_head	*list,
+	bool			do_discard)
 {
 	struct xfs_busy_extent	*busyp, *n;
 	struct xfs_perag	*pag = NULL;
@@ -2999,7 +3018,10 @@
 			agno = busyp->agno;
 		}
  
-		xfs_alloc_busy_clear_one(mp, pag, busyp);
+		if (do_discard && busyp->length)
+			busyp->flags = XFS_ALLOC_BUSY_DISCARDED;
+		else
+			xfs_alloc_busy_clear_one(mp, pag, busyp);
 	}
  
 	if (pag) {
@@ -140,7 +140,8 @@
 	xfs_agblock_t bno, xfs_extlen_t len);
  
 void
-xfs_alloc_busy_clear(struct xfs_mount *mp, struct list_head *list);
+xfs_alloc_busy_clear(struct xfs_mount *mp, struct list_head *list,
+	bool do_discard);
  
 int
 xfs_alloc_busy_search(struct xfs_mount *mp, xfs_agnumber_t agno,
@@ -29,6 +29,7 @@
 #include "xfs_mount.h"
 #include "xfs_error.h"
 #include "xfs_alloc.h"
+#include "xfs_discard.h"
  
 /*
  * Perform initial CIL structure initialisation. If the CIL is not
  
  
@@ -361,18 +362,28 @@
 	int	abort)
 {
 	struct xfs_cil_ctx	*ctx = args;
+	struct xfs_mount	*mp = ctx->cil->xc_log->l_mp;
  
 	xfs_trans_committed_bulk(ctx->cil->xc_log->l_ailp, ctx->lv_chain,
 					ctx->start_lsn, abort);
  
 	xfs_alloc_busy_sort(&ctx->busy_extents);
-	xfs_alloc_busy_clear(ctx->cil->xc_log->l_mp, &ctx->busy_extents);
+	xfs_alloc_busy_clear(mp, &ctx->busy_extents,
+			     (mp->m_flags & XFS_MOUNT_DISCARD) && !abort);
  
 	spin_lock(&ctx->cil->xc_cil_lock);
 	list_del(&ctx->committing);
 	spin_unlock(&ctx->cil->xc_cil_lock);
  
 	xlog_cil_free_logvec(ctx->lv_chain);
+
+	if (!list_empty(&ctx->busy_extents)) {
+		ASSERT(mp->m_flags & XFS_MOUNT_DISCARD);
+
+		xfs_discard_extents(mp, &ctx->busy_extents);
+		xfs_alloc_busy_clear(mp, &ctx->busy_extents, false);
+	}
+
 	kmem_free(ctx);
 }
  
@@ -224,6 +224,7 @@
 #define XFS_MOUNT_FS_SHUTDOWN	(1ULL << 4)	/* atomic stop of all filesystem
 						   operations, typically for
 						   disk errors in metadata */
+#define XFS_MOUNT_DISCARD	(1ULL << 5)	/* discard unused blocks */
 #define XFS_MOUNT_RETERR	(1ULL << 6)     /* return alignment errors to
 						   user */
 #define XFS_MOUNT_NOALIGN	(1ULL << 7)	/* turn off stripe alignment
@@ -609,7 +609,7 @@
 	struct xfs_trans	*tp)
 {
 	xfs_alloc_busy_sort(&tp->t_busy);
-	xfs_alloc_busy_clear(tp->t_mountp, &tp->t_busy);
+	xfs_alloc_busy_clear(tp->t_mountp, &tp->t_busy, false);
  
 	atomic_dec(&tp->t_mountp->m_active_trans);
 	xfs_trans_free_dqinfo(tp);
...	...	@@ -39,6 +39,12 @@
39	39	drive level write caching to be enabled, for devices that
40	40	support write barriers.
41	41
	42	+ discard
	43	+ Issue command to let the block device reclaim space freed by the
	44	+ filesystem. This is useful for SSD devices, thinly provisioned
	45	+ LUNs and virtual machine images, but may have a performance
	46	+ impact. This option is incompatible with the nodelaylog option.
	47	+
42	48	dmapi
43	49	Enable the DMAPI (Data Management API) event callouts.
44	50	Use with the "mtpt" option.
...	...	@@ -191,4 +191,33 @@
191	191	return -XFS_ERROR(EFAULT);
192	192	return 0;
193	193	}
	194	+
	195	+int
	196	+xfs_discard_extents(
	197	+ struct xfs_mount *mp,
	198	+ struct list_head *list)
	199	+{
	200	+ struct xfs_busy_extent *busyp;
	201	+ int error = 0;
	202	+
	203	+ list_for_each_entry(busyp, list, list) {
	204	+ trace_xfs_discard_extent(mp, busyp->agno, busyp->bno,
	205	+ busyp->length);
	206	+
	207	+ error = -blkdev_issue_discard(mp->m_ddev_targp->bt_bdev,
	208	+ XFS_AGB_TO_DADDR(mp, busyp->agno, busyp->bno),
	209	+ XFS_FSB_TO_BB(mp, busyp->length),
	210	+ GFP_NOFS, 0);
	211	+ if (error && error != EOPNOTSUPP) {
	212	+ xfs_info(mp,
	213	+ "discard failed for extent [0x%llu,%u], error %d",
	214	+ (unsigned long long)busyp->bno,
	215	+ busyp->length,
	216	+ error);
	217	+ return error;
	218	+ }
	219	+ }
	220	+
	221	+ return 0;
	222	+}
...	...	@@ -2,8 +2,10 @@
2	2	#define XFS_DISCARD_H 1
3	3
4	4	struct fstrim_range;
	5	+struct list_head;
5	6
6	7	extern int xfs_ioc_trim(struct xfs_mount , struct fstrim_range __user );
	8	+extern int xfs_discard_extents(struct xfs_mount , struct list_head );
7	9
8	10	#endif /* XFS_DISCARD_H */
...	...	@@ -110,8 +110,10 @@
110	110	#define MNTOPT_GQUOTANOENF "gqnoenforce"/* group quota limit enforcement */
111	111	#define MNTOPT_PQUOTANOENF "pqnoenforce"/* project quota limit enforcement */
112	112	#define MNTOPT_QUOTANOENF "qnoenforce" /* same as uqnoenforce */
113		-#define MNTOPT_DELAYLOG "delaylog" /* Delayed loging enabled */
114		-#define MNTOPT_NODELAYLOG "nodelaylog" /* Delayed loging disabled */
	113	+#define MNTOPT_DELAYLOG "delaylog" /* Delayed logging enabled */
	114	+#define MNTOPT_NODELAYLOG "nodelaylog" /* Delayed logging disabled */
	115	+#define MNTOPT_DISCARD "discard" /* Discard unused blocks */
	116	+#define MNTOPT_NODISCARD "nodiscard" /* Do not discard unused blocks */
115	117
116	118	/*
117	119	* Table driven mount option parser.
...	...	@@ -355,6 +357,10 @@
355	357	mp->m_flags \|= XFS_MOUNT_DELAYLOG;
356	358	} else if (!strcmp(this_char, MNTOPT_NODELAYLOG)) {
357	359	mp->m_flags &= ~XFS_MOUNT_DELAYLOG;
	360	+ } else if (!strcmp(this_char, MNTOPT_DISCARD)) {
	361	+ mp->m_flags \|= XFS_MOUNT_DISCARD;
	362	+ } else if (!strcmp(this_char, MNTOPT_NODISCARD)) {
	363	+ mp->m_flags &= ~XFS_MOUNT_DISCARD;
358	364	} else if (!strcmp(this_char, "ihashsize")) {
359	365	xfs_warn(mp,
360	366	"ihashsize no longer used, option is deprecated.");
...	...	@@ -388,6 +394,13 @@
388	394	return EINVAL;
389	395	}
390	396
	397	+ if ((mp->m_flags & XFS_MOUNT_DISCARD) &&
	398	+ !(mp->m_flags & XFS_MOUNT_DELAYLOG)) {
	399	+ xfs_warn(mp,
	400	+ "the discard option is incompatible with the nodelaylog option");
	401	+ return EINVAL;
	402	+ }
	403	+
391	404	#ifndef CONFIG_XFS_QUOTA
392	405	if (XFS_IS_QUOTA_RUNNING(mp)) {
393	406	xfs_warn(mp, "quota support not available in this kernel.");
...	...	@@ -488,6 +501,7 @@
488	501	{ XFS_MOUNT_FILESTREAMS, "," MNTOPT_FILESTREAM },
489	502	{ XFS_MOUNT_GRPID, "," MNTOPT_GRPID },
490	503	{ XFS_MOUNT_DELAYLOG, "," MNTOPT_DELAYLOG },
	504	+ { XFS_MOUNT_DISCARD, "," MNTOPT_DISCARD },
491	505	{ 0, NULL }
492	506	};
493	507	static struct proc_xfs_info xfs_info_unset[] = {
...	...	@@ -187,6 +187,8 @@
187	187	xfs_agnumber_t agno;
188	188	xfs_agblock_t bno;
189	189	xfs_extlen_t length;
	190	+ unsigned int flags;
	191	+#define XFS_ALLOC_BUSY_DISCARDED 0x01 /* undergoing a discard op. */
190	192	};
191	193
192	194	/*
...	...	@@ -2609,6 +2609,18 @@
2609	2609	xfs_agblock_t bend = bbno + busyp->length;
2610	2610
2611	2611	/*
	2612	+ * This extent is currently being discarded. Give the thread
	2613	+ * performing the discard a chance to mark the extent unbusy
	2614	+ * and retry.
	2615	+ */
	2616	+ if (busyp->flags & XFS_ALLOC_BUSY_DISCARDED) {
	2617	+ spin_unlock(&pag->pagb_lock);
	2618	+ delay(1);
	2619	+ spin_lock(&pag->pagb_lock);
	2620	+ return false;
	2621	+ }
	2622	+
	2623	+ /*
2612	2624	* If there is a busy extent overlapping a user allocation, we have
2613	2625	* no choice but to force the log and retry the search.
2614	2626	*
...	...	@@ -2813,7 +2825,8 @@
2813	2825	* If this is a metadata allocation, try to reuse the busy
2814	2826	* extent instead of trimming the allocation.
2815	2827	*/
2816		- if (!args->userdata) {
	2828	+ if (!args->userdata &&
	2829	+ !(busyp->flags & XFS_ALLOC_BUSY_DISCARDED)) {
2817	2830	if (!xfs_alloc_busy_update_extent(args->mp, args->pag,
2818	2831	busyp, fbno, flen,
2819	2832	false))
2820	2833
...	...	@@ -2979,10 +2992,16 @@
2979	2992	kmem_free(busyp);
2980	2993	}
2981	2994
	2995	+/*
	2996	+ * Remove all extents on the passed in list from the busy extents tree.
	2997	+ * If do_discard is set skip extents that need to be discarded, and mark
	2998	+ * these as undergoing a discard operation instead.
	2999	+ */
2982	3000	void
2983	3001	xfs_alloc_busy_clear(
2984	3002	struct xfs_mount *mp,
2985		- struct list_head *list)
	3003	+ struct list_head *list,
	3004	+ bool do_discard)
2986	3005	{
2987	3006	struct xfs_busy_extent busyp, n;
2988	3007	struct xfs_perag *pag = NULL;
...	...	@@ -2999,7 +3018,10 @@
2999	3018	agno = busyp->agno;
3000	3019	}
3001	3020
3002		- xfs_alloc_busy_clear_one(mp, pag, busyp);
	3021	+ if (do_discard && busyp->length)
	3022	+ busyp->flags = XFS_ALLOC_BUSY_DISCARDED;
	3023	+ else
	3024	+ xfs_alloc_busy_clear_one(mp, pag, busyp);
3003	3025	}
3004	3026
3005	3027	if (pag) {
...	...	@@ -140,7 +140,8 @@
140	140	xfs_agblock_t bno, xfs_extlen_t len);
141	141
142	142	void
143		-xfs_alloc_busy_clear(struct xfs_mount mp, struct list_head list);
	143	+xfs_alloc_busy_clear(struct xfs_mount mp, struct list_head list,
	144	+ bool do_discard);
144	145
145	146	int
146	147	xfs_alloc_busy_search(struct xfs_mount *mp, xfs_agnumber_t agno,
...	...	@@ -29,6 +29,7 @@
29	29	#include "xfs_mount.h"
30	30	#include "xfs_error.h"
31	31	#include "xfs_alloc.h"
	32	+#include "xfs_discard.h"
32	33
33	34	/*
34	35	* Perform initial CIL structure initialisation. If the CIL is not
35	36
36	37
...	...	@@ -361,18 +362,28 @@
361	362	int abort)
362	363	{
363	364	struct xfs_cil_ctx *ctx = args;
	365	+ struct xfs_mount *mp = ctx->cil->xc_log->l_mp;
364	366
365	367	xfs_trans_committed_bulk(ctx->cil->xc_log->l_ailp, ctx->lv_chain,
366	368	ctx->start_lsn, abort);
367	369
368	370	xfs_alloc_busy_sort(&ctx->busy_extents);
369		- xfs_alloc_busy_clear(ctx->cil->xc_log->l_mp, &ctx->busy_extents);
	371	+ xfs_alloc_busy_clear(mp, &ctx->busy_extents,
	372	+ (mp->m_flags & XFS_MOUNT_DISCARD) && !abort);
370	373
371	374	spin_lock(&ctx->cil->xc_cil_lock);
372	375	list_del(&ctx->committing);
373	376	spin_unlock(&ctx->cil->xc_cil_lock);
374	377
375	378	xlog_cil_free_logvec(ctx->lv_chain);
	379	+
	380	+ if (!list_empty(&ctx->busy_extents)) {
	381	+ ASSERT(mp->m_flags & XFS_MOUNT_DISCARD);
	382	+
	383	+ xfs_discard_extents(mp, &ctx->busy_extents);
	384	+ xfs_alloc_busy_clear(mp, &ctx->busy_extents, false);
	385	+ }
	386	+
376	387	kmem_free(ctx);
377	388	}
378	389
...	...	@@ -224,6 +224,7 @@
224	224	#define XFS_MOUNT_FS_SHUTDOWN (1ULL << 4) /* atomic stop of all filesystem
225	225	operations, typically for
226	226	disk errors in metadata */
	227	+#define XFS_MOUNT_DISCARD (1ULL << 5) /* discard unused blocks */
227	228	#define XFS_MOUNT_RETERR (1ULL << 6) /* return alignment errors to
228	229	user */
229	230	#define XFS_MOUNT_NOALIGN (1ULL << 7) /* turn off stripe alignment
...	...	@@ -609,7 +609,7 @@
609	609	struct xfs_trans *tp)
610	610	{
611	611	xfs_alloc_busy_sort(&tp->t_busy);
612		- xfs_alloc_busy_clear(tp->t_mountp, &tp->t_busy);
	612	+ xfs_alloc_busy_clear(tp->t_mountp, &tp->t_busy, false);
613	613
614	614	atomic_dec(&tp->t_mountp->m_active_trans);
615	615	xfs_trans_free_dqinfo(tp);