MFC r275568:

Count consecutive read requests as blocking in CTL for files and ZVOLs.

Technically read requests can be executed in any order or simultaneously
since they are not changing any data.  But ZFS prefetcher goes crasy when
it receives consecutive requests from different threads.  Since prefetcher
works on level of separate blocks, instead of two consecutive 128K requests
it may receive 32 8K requests in mixed order.

This patch is more workaround then a real fix, and it does not fix all of
prefetcher problems, but it improves sequential read speed by 3-4x times
in some configurations.  On the other side it may hurt performance if
some backing store has no prefetch, that is why it is disabled by default
for raw devices.
This commit is contained in:
mav 2014-12-18 08:46:53 +00:00
parent fe777840de
commit a0e66849d7
6 changed files with 75 additions and 21 deletions

View File

@ -435,7 +435,9 @@ static int ctl_inquiry_evpd_lbp(struct ctl_scsiio *ctsio, int alloc_len);
static int ctl_inquiry_evpd(struct ctl_scsiio *ctsio);
static int ctl_inquiry_std(struct ctl_scsiio *ctsio);
static int ctl_get_lba_len(union ctl_io *io, uint64_t *lba, uint64_t *len);
static ctl_action ctl_extent_check(union ctl_io *io1, union ctl_io *io2);
static ctl_action ctl_extent_check(union ctl_io *io1, union ctl_io *io2,
bool seq);
static ctl_action ctl_extent_check_seq(union ctl_io *io1, union ctl_io *io2);
static ctl_action ctl_check_for_blockage(struct ctl_lun *lun,
union ctl_io *pending_io, union ctl_io *ooa_io);
static ctl_action ctl_check_ooa(struct ctl_lun *lun, union ctl_io *pending_io,
@ -4592,6 +4594,17 @@ ctl_alloc_lun(struct ctl_softc *ctl_softc, struct ctl_lun *ctl_lun,
if (value != NULL && strcmp(value, "on") == 0)
lun->flags |= CTL_LUN_READONLY;
lun->serseq = CTL_LUN_SERSEQ_OFF;
if (be_lun->flags & CTL_LUN_FLAG_SERSEQ_READ)
lun->serseq = CTL_LUN_SERSEQ_READ;
value = ctl_get_opt(&be_lun->options, "serseq");
if (value != NULL && strcmp(value, "on") == 0)
lun->serseq = CTL_LUN_SERSEQ_ON;
else if (value != NULL && strcmp(value, "read") == 0)
lun->serseq = CTL_LUN_SERSEQ_READ;
else if (value != NULL && strcmp(value, "off") == 0)
lun->serseq = CTL_LUN_SERSEQ_OFF;
lun->ctl_softc = ctl_softc;
TAILQ_INIT(&lun->ooa_queue);
TAILQ_INIT(&lun->blocked_queue);
@ -10754,15 +10767,15 @@ ctl_get_lba_len(union ctl_io *io, uint64_t *lba, uint64_t *len)
}
static ctl_action
ctl_extent_check_lba(uint64_t lba1, uint64_t len1, uint64_t lba2, uint64_t len2)
ctl_extent_check_lba(uint64_t lba1, uint64_t len1, uint64_t lba2, uint64_t len2,
bool seq)
{
uint64_t endlba1, endlba2;
endlba1 = lba1 + len1 - 1;
endlba1 = lba1 + len1 - (seq ? 0 : 1);
endlba2 = lba2 + len2 - 1;
if ((endlba1 < lba2)
|| (endlba2 < lba1))
if ((endlba1 < lba2) || (endlba2 < lba1))
return (CTL_ACTION_PASS);
else
return (CTL_ACTION_BLOCK);
@ -10801,23 +10814,39 @@ ctl_extent_check_unmap(union ctl_io *io, uint64_t lba2, uint64_t len2)
}
static ctl_action
ctl_extent_check(union ctl_io *io1, union ctl_io *io2)
ctl_extent_check(union ctl_io *io1, union ctl_io *io2, bool seq)
{
uint64_t lba1, lba2;
uint64_t len1, len2;
int retval;
if (ctl_get_lba_len(io1, &lba1, &len1) != 0)
return (CTL_ACTION_ERROR);
retval = ctl_extent_check_unmap(io2, lba1, len1);
if (retval != CTL_ACTION_ERROR)
return (retval);
if (ctl_get_lba_len(io2, &lba2, &len2) != 0)
return (CTL_ACTION_ERROR);
return (ctl_extent_check_lba(lba1, len1, lba2, len2));
retval = ctl_extent_check_unmap(io1, lba2, len2);
if (retval != CTL_ACTION_ERROR)
return (retval);
if (ctl_get_lba_len(io1, &lba1, &len1) != 0)
return (CTL_ACTION_ERROR);
return (ctl_extent_check_lba(lba1, len1, lba2, len2, seq));
}
static ctl_action
ctl_extent_check_seq(union ctl_io *io1, union ctl_io *io2)
{
uint64_t lba1, lba2;
uint64_t len1, len2;
if (ctl_get_lba_len(io1, &lba1, &len1) != 0)
return (CTL_ACTION_ERROR);
if (ctl_get_lba_len(io2, &lba2, &len2) != 0)
return (CTL_ACTION_ERROR);
if (lba1 + len1 == lba2)
return (CTL_ACTION_BLOCK);
return (CTL_ACTION_PASS);
}
static ctl_action
@ -10906,12 +10935,18 @@ ctl_check_for_blockage(struct ctl_lun *lun, union ctl_io *pending_io,
case CTL_SER_BLOCK:
return (CTL_ACTION_BLOCK);
case CTL_SER_EXTENT:
return (ctl_extent_check(pending_io, ooa_io));
return (ctl_extent_check(ooa_io, pending_io,
(lun->serseq == CTL_LUN_SERSEQ_ON)));
case CTL_SER_EXTENTOPT:
if ((lun->mode_pages.control_page[CTL_PAGE_CURRENT].queue_flags
& SCP_QUEUE_ALG_MASK) != SCP_QUEUE_ALG_UNRESTRICTED)
return (ctl_extent_check(pending_io, ooa_io));
/* FALLTHROUGH */
return (ctl_extent_check(ooa_io, pending_io,
(lun->serseq == CTL_LUN_SERSEQ_ON)));
return (CTL_ACTION_PASS);
case CTL_SER_EXTENTSEQ:
if (lun->serseq != CTL_LUN_SERSEQ_OFF)
return (ctl_extent_check_seq(ooa_io, pending_io));
return (CTL_ACTION_PASS);
case CTL_SER_PASS:
return (CTL_ACTION_PASS);
case CTL_SER_BLOCKOPT:
@ -12442,7 +12477,7 @@ ctl_cmd_pattern_match(struct ctl_scsiio *ctsio, struct ctl_error_desc *desc)
return (CTL_LUN_PAT_NONE);
action = ctl_extent_check_lba(lba1, len1, desc->lba_range.lba,
desc->lba_range.len);
desc->lba_range.len, FALSE);
/*
* A "pass" means that the LBA ranges don't overlap, so
* this doesn't match the user's range criteria.

View File

@ -85,7 +85,8 @@ typedef enum {
CTL_LUN_FLAG_DEVID = 0x20,
CTL_LUN_FLAG_DEV_TYPE = 0x40,
CTL_LUN_FLAG_UNMAP = 0x80,
CTL_LUN_FLAG_OFFLINE = 0x100
CTL_LUN_FLAG_OFFLINE = 0x100,
CTL_LUN_FLAG_SERSEQ_READ = 0x200
} ctl_backend_lun_flags;
#ifdef _KERNEL

View File

@ -2207,6 +2207,8 @@ ctl_be_block_create(struct ctl_be_block_softc *softc, struct ctl_lun_req *req)
be_lun->ctl_be_lun.flags |= CTL_LUN_FLAG_OFFLINE;
if (unmap)
be_lun->ctl_be_lun.flags |= CTL_LUN_FLAG_UNMAP;
if (be_lun->dispatch != ctl_be_block_dispatch_dev)
be_lun->ctl_be_lun.flags |= CTL_LUN_FLAG_SERSEQ_READ;
be_lun->ctl_be_lun.be_lun = be_lun;
be_lun->ctl_be_lun.maxlba = (be_lun->size_blocks == 0) ?
0 : (be_lun->size_blocks - 1);

View File

@ -97,6 +97,7 @@ typedef enum {
CTL_SER_BLOCKOPT,
CTL_SER_EXTENT,
CTL_SER_EXTENTOPT,
CTL_SER_EXTENTSEQ,
CTL_SER_PASS,
CTL_SER_SKIP
} ctl_serialize_action;
@ -182,6 +183,12 @@ typedef enum {
CTL_LUN_READONLY = 0x800
} ctl_lun_flags;
typedef enum {
CTL_LUN_SERSEQ_OFF,
CTL_LUN_SERSEQ_READ,
CTL_LUN_SERSEQ_ON
} ctl_lun_serseq;
typedef enum {
CTLBLOCK_FLAG_NONE = 0x00,
CTLBLOCK_FLAG_INVALID = 0x01
@ -386,6 +393,7 @@ struct ctl_lun {
struct ctl_id target;
uint64_t lun;
ctl_lun_flags flags;
ctl_lun_serseq serseq;
STAILQ_HEAD(,ctl_error_desc) error_list;
uint64_t error_serial;
struct ctl_softc *ctl_softc;

View File

@ -59,12 +59,13 @@
#define bO CTL_SER_BLOCKOPT /* Optional block */
#define xT CTL_SER_EXTENT /* Extent check */
#define xO CTL_SER_EXTENTOPT /* Optional extent check */
#define xS CTL_SER_EXTENTSEQ /* Sequential extent check */
static ctl_serialize_action
ctl_serialize_table[CTL_SERIDX_COUNT][CTL_SERIDX_COUNT] = {
/**>IDX_ :: 2nd:TUR RD WRT UNM MDSN MDSL RQSN INQ RDCP RES LSNS FMT STR*/
/*TUR */{ pS, pS, pS, pS, bK, bK, bK, pS, pS, bK, pS, bK, bK},
/*READ */{ pS, pS, xT, bO, bK, bK, bK, pS, pS, bK, pS, bK, bK},
/*READ */{ pS, xS, xT, bO, bK, bK, bK, pS, pS, bK, pS, bK, bK},
/*WRITE */{ pS, xT, xT, bO, bK, bK, bK, pS, pS, bK, pS, bK, bK},
/*UNMAP */{ pS, xO, xO, pS, bK, bK, bK, pS, pS, bK, pS, bK, bK},
/*MD_SNS */{ bK, bK, bK, bK, pS, bK, bK, pS, pS, bK, pS, bK, bK},

View File

@ -34,7 +34,7 @@
.\" $Id: //depot/users/kenm/FreeBSD-test2/usr.sbin/ctladm/ctladm.8#3 $
.\" $FreeBSD$
.\"
.Dd November 5, 2014
.Dd December 6, 2014
.Dt CTLADM 8
.Os
.Sh NAME
@ -995,6 +995,13 @@ command sequence order shall be explicitly handled by the application
client through the selection of appropriate commands and task attributes.
The default value is "restricted". It improves data integrity, but may
introduce some additional delays.
.It Va serseq
Set to "on" to serialize conseсutive reads/writes.
Set to "read" to serialize conseсutive reads.
Set to "off" to allow them be issued in parallel.
Parallel issue of consecutive operations may confuse logic of the
backing file system, hurting performance; but it may improve performance
of backing stores without prefetch/write-back.
.It Va rpm
Specifies medium rotation rate of the device: 0 -- not reported,
1 -- non-rotating (SSD), >1024 -- value in revolutions per minute.