scsi: target: XCOPY performance
Subsystems (2)
cover scsi: target: XCOPY performance
These changes remove unnecessary heap allocations in the XCOPY READ/WRITE dispatch loop.
Synthetic benchmarks on my laptop using the libiscsi iscsi-dd utility (--xcopy --max 1 --blocks 65535 src=dst) against a target backed by an 8G zram (DEBUG_KMEMLEAK=y) iblock backstore (avg across four runs) show:
before: 5.35776G/s after: 6.12636G/s (approx. +14%)
Feedback appreciated.
Cheers, David
----------------------------------------------------------------
David Disseldorp (5):
scsi: target: use #def for xcopy descriptor len
scsi: target: drop xcopy DISK BLOCK LENGTH debug
scsi: target: avoid per-loop XCOPY buffer allocations
scsi: target: increase XCOPY I/O size
scsi: target: avoid XCOPY per-loop read/write cmd allocations
drivers/target/target_core_xcopy.c | 135 ++++++++++------------------- drivers/target/target_core_xcopy.h | 14 +-- 2 files changed, 57 insertions(+), 92 deletions(-)
RFC PATCH 1/5 scsi: target: use #def for xcopy descriptor len
--- drivers/target/target_core_xcopy.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/target/target_core_xcopy.c +1 -1
@@ -134,7 +134,7 @@ static int target_xcopy_parse_tiddesc_e4(struct se_cmd *se_cmd, struct xcopy_op* Assigned designator */ desig_len = desc[7 ]; -if (desig_len !=16 ) { +if (desig_len !=XCOPY_NAA_IEEE_REGEX_LEN ) {pr_err ("XCOPY 0xe4: invalid desig_len: %d\n" , (int )desig_len);return -EINVAL ; }
Signature · 1 line
2.16.4
Re: [RFC PATCH 1/5] scsi: target: use #def for xcopy descriptor len
Quoted 1 line · David Disseldorp
Signed-off-by David Disseldorp <[email protected]>
Looks good,
RFC PATCH 2/5 scsi: target: drop xcopy DISK BLOCK LENGTH debug
In reply to David DisseldorpThe DISK BLOCK LENGTH field is carried with XCOPY target descriptors on the wire, but is currently unmarshalled during 0x02 segment descriptor passing. The unmarshalled value is currently unused, so drop it.
--- drivers/target/target_core_xcopy.c | 5 ----- drivers/target/target_core_xcopy.h | 1 - 2 files changed, 6 deletions(-)
drivers/target/target_core_xcopy.c +0 -5
@@ -315,11 +315,6 @@ static int target_xcopy_parse_segdesc_02(struct se_cmd *se_cmd, struct xcopy_op xop->nolb, (unsigned long long )xop->src_lba, (unsigned long long)xop->dst_lba); -if (dc !=0 ) { - xop->dbl =get_unaligned_be24 (&desc[29 ]); - -pr_debug ("XCOPY seg desc 0x02: DC=1 w/ dbl: %u\n" , xop->dbl); - }return 0 ; }
drivers/target/target_core_xcopy.h +0 -1
@@ -35,7 +35,6 @@ struct xcopy_op {unsigned short stdi;unsigned short dtdi;unsigned short nolb; -unsigned int dbl;struct xcopy_pt_cmd *src_pt_cmd;struct xcopy_pt_cmd *dst_pt_cmd;
Signature · 1 line
2.16.4
Re: [RFC PATCH 2/5] scsi: target: drop xcopy DISK BLOCK LENGTH debug
Quoted 3 lines · David Disseldorp
The DISK BLOCK LENGTH field is carried with XCOPY target descriptors on the wire, but is currently unmarshalled during
0x02segment descriptor passing. The unmarshalled value is currently unused, so drop it.
Looks good,
RFC PATCH 3/5 scsi: target: avoid per-loop XCOPY buffer allocations
In reply to David DisseldorpThe main target_xcopy_do_work() loop unnecessarily allocates an I/O buffer with each synchronous READ / WRITE pair. This commit significantly reduces allocations by reusing the XCOPY I/O buffer when possible.
--- drivers/target/target_core_xcopy.c | 95 +++++++++++++++----------------------- drivers/target/target_core_xcopy.h | 1 + 2 files changed, 39 insertions(+), 57 deletions(-)
drivers/target/target_core_xcopy.c +38 -57
@@ -499,7 +499,6 @@ void target_xcopy_release_pt(void)* @cdb: SCSI CDB to be copied into @xpt_cmd. * @remote_port: If false, use the LUN through which the XCOPY command has * been received. If true, use @se_dev->xcopy_lun. -* @alloc_mem: Whether or not to allocate an SGL list. * * Set up a SCSI command (READ or WRITE) that will be used to execute an * XCOPY command.
@@ -509,12 +508,11 @@ static int target_xcopy_setup_pt_cmd(struct xcopy_op *xop, struct se_device *se_dev, unsigned char *cdb, - bool remote_port, - bool alloc_mem) + bool remote_port) { struct se_cmd *cmd = &xpt_cmd->se_cmd;sense_reason_t sense_rc; -int ret =0 , rc; +int ret =0 ;/* * Setup LUN+port to honor reservations based upon xop->op_origin for
@@ -536,36 +534,21 @@ static int target_xcopy_setup_pt_cmd( goto out; } -if (alloc_mem) { - rc =target_alloc_sgl (&cmd->t_data_sg, &cmd->t_data_nents, - cmd->data_length, false, false); -if (rc <0 ) { - ret = rc; - goto out; - } -/* -* Set this bit so that transport_free_pages() allows the -* caller to release SGLs + physical memory allocated by -* transport_generic_get_mem().. -*/ - cmd->se_cmd_flags |=SCF_PASSTHROUGH_SG_TO_MEM_NOALLOC ; - }else { -/* -* Here the previously allocated SGLs for the internal READ -* are mapped zero-copy to the internal WRITE. -*/ - sense_rc =transport_generic_map_mem_to_cmd (cmd, - xop->xop_data_sg, xop->xop_data_nents, -NULL ,0 ); -if (sense_rc) { - ret = -EINVAL ; - goto out; - } - -pr_debug ("Setup PASSTHROUGH_NOALLOC t_data_sg: %p t_data_nents:" -" %u\n" , cmd->t_data_sg, cmd->t_data_nents); +/* +* Here the previously allocated SGLs for the XCOPY are mapped zero-copy +* to the internal READ or WRITE. +*/ + sense_rc =transport_generic_map_mem_to_cmd (cmd, + xop->xop_data_sg, xop->xop_data_nents, +NULL ,0 ); +if (sense_rc) { + ret = -EINVAL ; + goto out; } +pr_debug ("Setup PASSTHROUGH_NOALLOC t_data_sg: %p t_data_nents:" +" %u\n" , cmd->t_data_sg, cmd->t_data_nents); +return 0 ; out:
@@ -626,15 +609,13 @@ static int target_xcopy_read_source( xop->src_pt_cmd = xpt_cmd; rc =target_xcopy_setup_pt_cmd (xpt_cmd, xop, src_dev, &cdb[0 ], - remote_port, true); + remote_port);if (rc <0 ) { ec_cmd->scsi_status = xpt_cmd->se_cmd.scsi_status;transport_generic_free_cmd (se_cmd,0 );return rc; } - xop->xop_data_sg = se_cmd->t_data_sg; - xop->xop_data_nents = se_cmd->t_data_nents;pr_debug ("XCOPY-READ: Saved xop->xop_data_sg: %p, num: %u for READ" " memory\n" , xop->xop_data_sg, xop->xop_data_nents);
@@ -644,12 +625,6 @@ static int target_xcopy_read_source(transport_generic_free_cmd (se_cmd,0 );return rc; } -/* -* Clear off the allocated t_data_sg, that has been saved for -* zero-copy WRITE submission reuse in struct xcopy_op.. -*/ - se_cmd->t_data_sg =NULL ; - se_cmd->t_data_nents =0 ;return 0 ; }
@@ -688,19 +663,9 @@ static int target_xcopy_write_destination( xop->dst_pt_cmd = xpt_cmd; rc =target_xcopy_setup_pt_cmd (xpt_cmd, xop, dst_dev, &cdb[0 ], - remote_port, false); + remote_port);if (rc <0 ) { -struct se_cmd *src_cmd = &xop->src_pt_cmd->se_cmd; ec_cmd->scsi_status = xpt_cmd->se_cmd.scsi_status; -/* -* If the failure happened before the t_mem_list hand-off in -* target_xcopy_setup_pt_cmd(), Reset memory + clear flag so that -* core releases this memory on error during X-COPY WRITE I/O. -*/ - src_cmd->se_cmd_flags &= ~SCF_PASSTHROUGH_SG_TO_MEM_NOALLOC ; - src_cmd->t_data_sg = xop->xop_data_sg; - src_cmd->t_data_nents = xop->xop_data_nents; -transport_generic_free_cmd (se_cmd,0 );return rc; }
@@ -708,7 +673,6 @@ static int target_xcopy_write_destination( rc =target_xcopy_issue_pt_cmd (xpt_cmd);if (rc <0 ) { ec_cmd->scsi_status = xpt_cmd->se_cmd.scsi_status; - se_cmd->se_cmd_flags &= ~SCF_PASSTHROUGH_SG_TO_MEM_NOALLOC ;transport_generic_free_cmd (se_cmd,0 );return rc; }
@@ -724,7 +688,7 @@ static void target_xcopy_do_work(struct work_struct *work)sector_t src_lba, dst_lba, end_lba;unsigned int max_sectors;int rc =0 ; -unsigned short nolb, cur_nolb, max_nolb, copied_nolb =0 ; +unsigned short nolb, max_nolb, copied_nolb =0 ;if (target_parse_xcopy_cmd (xop) !=TCM_NO_SENSE ) goto err_free;
@@ -754,7 +718,24 @@ static void target_xcopy_do_work(struct work_struct *work) (unsigned long long)src_lba, (unsigned long long)dst_lba);while (src_lba < end_lba) { - cur_nolb =min (nolb, max_nolb); +unsigned short cur_nolb =min (nolb, max_nolb); +u32 cur_bytes = cur_nolb * src_dev->dev_attrib.block_size; + +if (cur_bytes != xop->xop_data_bytes) { +/* +* (Re)allocate a buffer large enough to hold the XCOPY +* I/O size, which can be reused each read / write loop. +*/ +target_free_sgl (xop->xop_data_sg, xop->xop_data_nents); + rc =target_alloc_sgl (&xop->xop_data_sg, + &xop->xop_data_nents, + cur_bytes, + false, false); +if (rc <0 ) { + goto out; + } + xop->xop_data_bytes = cur_bytes; + }pr_debug ("target_xcopy_do_work: Calling read src_dev: %p src_lba: %llu," " cur_nolb: %hu\n" , src_dev, (unsigned long long )src_lba, cur_nolb);
@@ -785,12 +766,11 @@ static void target_xcopy_do_work(struct work_struct *work) nolb -= cur_nolb;transport_generic_free_cmd (&xop->src_pt_cmd->se_cmd,0 ); - xop->dst_pt_cmd->se_cmd.se_cmd_flags &= ~SCF_PASSTHROUGH_SG_TO_MEM_NOALLOC ; -transport_generic_free_cmd (&xop->dst_pt_cmd->se_cmd,0 ); }xcopy_pt_undepend_remotedev (xop); +target_free_sgl (xop->xop_data_sg, xop->xop_data_nents);kfree (xop); pr_debug("target_xcopy_do_work: Final src_lba: %llu, dst_lba: %llu\n",
@@ -804,6 +784,7 @@ static void target_xcopy_do_work(struct work_struct *work) out:xcopy_pt_undepend_remotedev (xop); +target_free_sgl (xop->xop_data_sg, xop->xop_data_nents); err_free:kfree (xop);
drivers/target/target_core_xcopy.h +1 -0
@@ -39,6 +39,7 @@ struct xcopy_op {struct xcopy_pt_cmd *src_pt_cmd;struct xcopy_pt_cmd *dst_pt_cmd; +u32 xop_data_bytes;u32 xop_data_nents;struct scatterlist *xop_data_sg;struct work_struct xop_work;
Signature · 1 line
2.16.4
Re: [RFC PATCH 3/5] scsi: target: avoid per-loop XCOPY buffer allocations
drivers/target/target_core_xcopy.c static int target_xcopy_setup_pt_cmd( · line 540
Quoted 4 lines
drivers/target/target_core_xcopy.c +4
-0+/* +* Here the previously allocated SGLs for the XCOPY are mapped zero-copy +* to the internal READ or WRITE. +*/
This comments is rather pointless and a little confusing. I'd use the chance to kill it rather than bringing it over from the previous code.
drivers/target/target_core_xcopy.c static int target_xcopy_setup_pt_cmd( · line 547
Quoted 7 lines
drivers/target/target_core_xcopy.c +6
-0+ sense_rc =transport_generic_map_mem_to_cmd (cmd, + xop->xop_data_sg, xop->xop_data_nents, +NULL ,0 ); +if (sense_rc) { + ret = -EINVAL ; + goto out; }
This could simply be:
if (transport_generic_map_mem_to_cmd(cmd, xop->xop_data_sg, xop->xop_data_nents, NULL, 0)) return -EINVAL;
drivers/target/target_core_xcopy.c static void target_xcopy_do_work(struct work_struct *work) · line 736
Quoted 7 lines
drivers/target/target_core_xcopy.c +7
-0+ rc =target_alloc_sgl (&xop->xop_data_sg, + &xop->xop_data_nents, + cur_bytes, + false, false); +if (rc <0 ) { + goto out; + }
rc = target_alloc_sgl(&xop->xop_data_sg, &xop->xop_data_nents, cur_bytes, false, false); if (rc < 0) goto out;
Otherwise looks good:
RFC PATCH 4/5 scsi: target: increase XCOPY I/O size
In reply to David DisseldorpThe I/O size is already bound by dev_attrib.hw_max_sectors, so increase the hardcoded XCOPY_MAX_SECTORS maximum to improve performance against backstores with high-latency.
--- drivers/target/target_core_xcopy.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/target/target_core_xcopy.h +1 -1
@@ -5,7 +5,7 @@#define XCOPY_TARGET_DESC_LEN 32#define XCOPY_SEGMENT_DESC_LEN 28#define XCOPY_NAA_IEEE_REGEX_LEN 16 -#define XCOPY_MAX_SECTORS 1024 +#define XCOPY_MAX_SECTORS 4096/* * SPC4r37 6.4.6.1
Signature · 1 line
2.16.4
Re: [RFC PATCH 4/5] scsi: target: increase XCOPY I/O size
Quoted 3 lines · David Disseldorp
The I/O size is already bound by dev_attrib.hw_max_sectors, so increase the hardcoded
XCOPY_MAX_SECTORSmaximum to improve performance against backstores with high-latency.
Looks good,
RFC PATCH 5/5 scsi: target: avoid XCOPY per-loop read/write cmd allocations
In reply to David DisseldorpReads and writes in the XCOPY loop are synchronous, so needn't be allocated / freed with each loop.
--- drivers/target/target_core_xcopy.c | 33 +++++++++------------------------ drivers/target/target_core_xcopy.h | 10 +++++++--- 2 files changed, 16 insertions(+), 27 deletions(-)
drivers/target/target_core_xcopy.c +9 -24
@@ -374,12 +374,6 @@ static int target_xcopy_parse_segment_descriptors(struct se_cmd *se_cmd,* Start xcopy_pt ops */ -struct xcopy_pt_cmd { -struct se_cmd se_cmd; -struct completion xpt_passthrough_sem; -unsigned char sense_buffer[TRANSPORT_SENSE_BUFFER ]; -}; -struct se_portal_group xcopy_pt_tpg;static struct se_session xcopy_pt_sess;static struct se_node_acl xcopy_pt_nacl;
@@ -410,7 +404,8 @@ static void xcopy_pt_release_cmd(struct se_cmd *se_cmd)struct xcopy_pt_cmd *xpt_cmd =container_of (se_cmd, struct xcopy_pt_cmd, se_cmd); -kfree (xpt_cmd); +/* xpt_cmd is backed by parent xop, nothing to free */ +pr_debug ("xpt_cmd done: %p\n" , xpt_cmd); }static int xcopy_pt_check_stop_free (struct se_cmd *se_cmd)
@@ -582,18 +577,14 @@ static int target_xcopy_read_source( sector_t src_lba, u32 src_sectors) { -struct xcopy_pt_cmd *xpt_cmd; +struct xcopy_pt_cmd *xpt_cmd = &xop->src_pt_cmd;struct se_cmd *se_cmd;u32 length = (src_sectors * src_dev->dev_attrib.block_size);int rc;unsigned char cdb[16 ];bool remote_port = (xop->op_origin =XCOL_DEST_RECV_OP ); - xpt_cmd =kzalloc (sizeof (struct xcopy_pt_cmd ),GFP_KERNEL ); -if (!xpt_cmd) { -pr_err ("Unable to allocate xcopy_pt_cmd\n" ); -return -ENOMEM ; - } +memset (xpt_cmd,0 ,sizeof (*xpt_cmd));init_completion (&xpt_cmd->xpt_passthrough_sem); se_cmd = &xpt_cmd->se_cmd;
@@ -606,7 +597,6 @@ static int target_xcopy_read_source(transport_init_se_cmd (se_cmd, &xcopy_pt_tfo, &xcopy_pt_sess, length,DMA_FROM_DEVICE ,0 , &xpt_cmd->sense_buffer[0 ]); - xop->src_pt_cmd = xpt_cmd; rc =target_xcopy_setup_pt_cmd (xpt_cmd, xop, src_dev, &cdb[0 ], remote_port);
@@ -636,18 +626,14 @@ static int target_xcopy_write_destination( sector_t dst_lba, u32 dst_sectors) { -struct xcopy_pt_cmd *xpt_cmd; +struct xcopy_pt_cmd *xpt_cmd = &xop->dst_pt_cmd;struct se_cmd *se_cmd;u32 length = (dst_sectors * dst_dev->dev_attrib.block_size);int rc;unsigned char cdb[16 ];bool remote_port = (xop->op_origin =XCOL_SOURCE_RECV_OP ); - xpt_cmd =kzalloc (sizeof (struct xcopy_pt_cmd ),GFP_KERNEL ); -if (!xpt_cmd) { -pr_err ("Unable to allocate xcopy_pt_cmd\n" ); -return -ENOMEM ; - } +memset (xpt_cmd,0 ,sizeof (*xpt_cmd));init_completion (&xpt_cmd->xpt_passthrough_sem); se_cmd = &xpt_cmd->se_cmd;
@@ -660,7 +646,6 @@ static int target_xcopy_write_destination(transport_init_se_cmd (se_cmd, &xcopy_pt_tfo, &xcopy_pt_sess, length,DMA_TO_DEVICE ,0 , &xpt_cmd->sense_buffer[0 ]); - xop->dst_pt_cmd = xpt_cmd; rc =target_xcopy_setup_pt_cmd (xpt_cmd, xop, dst_dev, &cdb[0 ], remote_port);
@@ -754,7 +739,7 @@ static void target_xcopy_do_work(struct work_struct *work) rc =target_xcopy_write_destination (ec_cmd, xop, dst_dev, dst_lba, cur_nolb);if (rc <0 ) { -transport_generic_free_cmd (&xop->src_pt_cmd->se_cmd,0 ); +transport_generic_free_cmd (&xop->src_pt_cmd.se_cmd,0 ); goto out; }
@@ -765,8 +750,8 @@ static void target_xcopy_do_work(struct work_struct *work) copied_nolb += cur_nolb; nolb -= cur_nolb; -transport_generic_free_cmd (&xop->src_pt_cmd->se_cmd,0 ); -transport_generic_free_cmd (&xop->dst_pt_cmd->se_cmd,0 ); +transport_generic_free_cmd (&xop->src_pt_cmd.se_cmd,0 ); +transport_generic_free_cmd (&xop->dst_pt_cmd.se_cmd,0 ); }xcopy_pt_undepend_remotedev (xop);
drivers/target/target_core_xcopy.h +7 -3
@@ -18,7 +18,11 @@ enum xcopy_origin_list { XCOL_DEST_RECV_OP = 0x02, }; -struct xcopy_pt_cmd ; +struct xcopy_pt_cmd { +struct se_cmd se_cmd; +struct completion xpt_passthrough_sem; +unsigned char sense_buffer[TRANSPORT_SENSE_BUFFER ]; +};struct xcopy_op {int op_origin;
@@ -36,8 +40,8 @@ struct xcopy_op {unsigned short dtdi;unsigned short nolb; -struct xcopy_pt_cmd *src_pt_cmd; -struct xcopy_pt_cmd *dst_pt_cmd; +struct xcopy_pt_cmd src_pt_cmd; +struct xcopy_pt_cmd dst_pt_cmd;u32 xop_data_bytes;u32 xop_data_nents;
Signature · 1 line
2.16.4
Re: [RFC PATCH 5/5] scsi: target: avoid XCOPY per-loop read/write cmd allocations
Quoted 2 lines · David Disseldorp
Reads and writes in the XCOPY loop are synchronous, so needn't be allocated / freed with each loop.
That is true, but I think with your previous cleanups we can easily go one step further and just allocate a single command and sense buffer directly in struct xcopy_op, and just have local completions on the stack.
Re: [RFC PATCH 5/5] scsi: target: avoid XCOPY per-loop read/write cmd allocations
Hi Christoph,
Thanks for the feedback...
Quoted 8 lines · Christoph Hellwig
Quoted 2 lines · David Disseldorp
Reads and writes in the XCOPY loop are synchronous, so needn't be allocated / freed with each loop.
That is true, but I think with your previous cleanups we can easily go one step further and just allocate a single command and sense buffer directly in struct
xcopy_op, and just have local completions on the stack.
I'm probably missing something, but having the (stack) completion separate to the se_cmd and sense buffer would mean that it's no longer a straightforward container_of() in the .check_stop_free() callback.
I've reworked this patch to put the entire xcopy_pt_cmd on the stack and will send it out with v2.
Cheers, David