Max Queue depth of HBA limited to 256 ?
Kenneth D. Merry
ken at freebsd.org
Tue Feb 5 21:18:39 UTC 2013
I'm able to get more than 255 commands outstanding to the controller in my
configuration.
For example:
dev.mps.0.%desc: LSI SAS2116
dev.mps.0.%driver: mps
dev.mps.0.%location: slot=6 function=0 handle=\_SB_.PCI0.S30_
dev.mps.0.%pnpinfo: vendor=0x1000 device=0x0064 subvendor=0x1000
subdevice=0x30c0 class=0x010700
dev.mps.0.%parent: pci0
dev.mps.0.debug_level: 4
dev.mps.0.disable_msix: 0
dev.mps.0.disable_msi: 0
dev.mps.0.firmware_version: 13.00.01.00
dev.mps.0.driver_version: 14.00.00.01-fbsd
dev.mps.0.io_cmds_active: 442
dev.mps.0.io_cmds_highwater: 464
dev.mps.0.chain_free: 354
dev.mps.0.chain_free_lowwater: 181
dev.mps.0.max_chains: 2048
dev.mps.0.chain_alloc_fail: 0
io_cmds_highwater is 464. Can you get more than 255 commands outstanding
if you use more than 1 target?
This is with 272 'dd' processes doing 1MB reads to 16 2TB and 3TB SAS
drives behind 2 3Gb Maxim expanders:
<SEAGATE ST32000444SS 0006> at scbus2 target 144 lun 0 (pass4,sg4,da0)
<SEAGATE ST32000444SS 0006> at scbus2 target 145 lun 0 (pass5,sg5,da1)
<SEAGATE ST33000650SS 0003> at scbus2 target 146 lun 0 (pass6,sg6,da2)
<SEAGATE ST33000650SS 0003> at scbus2 target 147 lun 0 (pass7,sg7,da3)
<SEAGATE ST32000444SS 0006> at scbus2 target 148 lun 0 (pass8,sg8,da4)
<SEAGATE ST32000444SS 0006> at scbus2 target 149 lun 0 (pass9,sg9,da5)
<SEAGATE ST33000650SS 0003> at scbus2 target 150 lun 0 (pass10,sg10,da6)
<SEAGATE ST33000650SS 0003> at scbus2 target 151 lun 0 (pass11,sg11,da7)
<SEAGATE ST32000444SS 0006> at scbus2 target 152 lun 0 (pass12,sg12,da8)
<SEAGATE ST32000444SS 0006> at scbus2 target 153 lun 0 (pass13,sg13,da9)
<SEAGATE ST32000444SS 0006> at scbus2 target 154 lun 0 (pass14,sg14,da10)
<SEAGATE ST32000444SS 0006> at scbus2 target 155 lun 0 (pass15,sg15,da11)
<SEAGATE ST33000650SS 0003> at scbus2 target 156 lun 0 (pass16,sg16,da12)
<SEAGATE ST33000650SS 0003> at scbus2 target 157 lun 0 (pass17,sg17,da13)
<SEAGATE ST33000650SS 0003> at scbus2 target 158 lun 0 (pass18,sg18,da14)
<SEAGATE ST33000650SS 0003> at scbus2 target 159 lun 0 (pass19,sg19,da15)
i.e. 17 iterations of this:
((i=0)); while [ $i -le 15 ]; do dd if=/dev/da$i of=/dev/null bs=1m & ((i++)); done
The individual drives see varying numbers of tags, but nowhere near the
maximum:
[root at storage-domain ~]# camcontrol tags da15 -v
(pass19:mps0:0:159:0): dev_openings 230
(pass19:mps0:0:159:0): dev_active 25
(pass19:mps0:0:159:0): devq_openings 230
(pass19:mps0:0:159:0): devq_queued 0
(pass19:mps0:0:159:0): held 0
(pass19:mps0:0:159:0): mintags 2
(pass19:mps0:0:159:0): maxtags 255
What kind of drive is the target?
Ken
On Wed, Jan 23, 2013 at 00:44:31 +0530, Desai, Kashyap wrote:
> LSI h/w needs more outstanding command in FW to get better Perf counts compare to other OS.
>
> Please suggest if whatever I have been observed is limitation from FreeBSD or we can tune it in Driver ?
> My goals is to pump ~1000 outstanding IOs to the HBA. I see that it never goes beyond 255.
>
> Thanks,
> Kashyap
>
> > -----Original Message-----
> > From: owner-freebsd-scsi at freebsd.org [mailto:owner-freebsd-
> > scsi at freebsd.org] On Behalf Of Desai, Kashyap
> > Sent: Monday, January 21, 2013 11:18 PM
> > To: Kenneth D. Merry
> > Cc: freebsd-scsi at freebsd.org; jhb at freebsd.org; McConnell, Stephen
> > Subject: RE: Max Queue depth of HBA limited to 256 ?
> >
> >
> >
> > > -----Original Message-----
> > > From: Kenneth D. Merry [mailto:ken at freebsd.org]
> > > Sent: Monday, January 21, 2013 10:35 PM
> > > To: Desai, Kashyap
> > > Cc: freebsd-scsi at freebsd.org; McConnell, Stephen; Saxena, Sumit;
> > > jhb at freebsd.org
> > > Subject: Re: Max Queue depth of HBA limited to 256 ?
> > >
> > > On Mon, Jan 21, 2013 at 20:15:47 +0530, Desai, Kashyap wrote:
> > > > Hi,
> > > >
> > > > I was trying to check few things on LSI controller, where we have
> > > > more
> > > than 256 queue depth support.
> > > > I added default maxtags in scsi/scsi_xpt.c as below. (Because I
> > > > don't
> > > want mattags to restrict any outstanding commands the LSI HBA.
> > > >
> > > > {
> > > > /* Default tagged queuing parameters for all devices */
> > > > {
> > > > T_ANY, SIP_MEDIA_REMOVABLE|SIP_MEDIA_FIXED,
> > > > /*vendor*/"*", /*product*/"*", /*revision*/"*"
> > > > },
> > > > /*quirks*/0, /*mintags*/2, /*maxtags*/1024 <--- Default
> > > maxtags were 256. I increase it to 10234
> > > > },
> > > >
> > > >
> > > > LSI's SAS-HBA and MR-HBA can support more than 256 outstanding
> > > commands in Firmware. But due to some reason, I am not able to pump
> > > more than 256 outstanding commands to the HBA.
> > > >
> > > > I used "rawio -p 256 /dev/da1" and more /dev/dax in loop. I have
> > > sysctl parameter in Driver to display outstanding "FW commands". Max
> > > value for FW outstanding only goes up to 256.
> > > >
> > > > Also from some other mail thread Subject "mfi driver performance", I
> > > found that folks talk about tuning queue depth _but_ nobody discussed
> > > to increase it beyond 256. Is there any limitation in FreeBSD ?
> > > >
> > >
> > > As Jim pointed out, one thing to check is the values passed into
> > > cam_sim_alloc(). In the case of the mps(4) driver, the calculation is
> > > in mps_attach():
> > >
> > > sc->num_reqs = MIN(MPS_REQ_FRAMES, sc->facts->RequestCredit);
> > >
> > > What is reported for the RequestCredit on this particular adapter?
> > >
> > > The other question is, what does 'camcontrol tags daX -v' show when
> > > you are running the test?
> >
> > Below is output of camcontrol tags da1 -v.
> >
> > dhcp-135-24-192-127# camcontrol tags da13 -v
> > (pass13:mrsas0:0:13:0): dev_openings 1024
> > (pass13:mrsas0:0:13:0): dev_active 0
> > (pass13:mrsas0:0:13:0): devq_openings 1024
> > (pass13:mrsas0:0:13:0): devq_queued 0
> > (pass13:mrsas0:0:13:0): held 0
> > (pass13:mrsas0:0:13:0): mintags 2
> > (pass13:mrsas0:0:13:0): maxtags 1024
> > dhcp-135-24-192-127# camcontrol tags da1 -v
> > (pass1:mrsas0:0:1:0): dev_openings 1024
> > (pass1:mrsas0:0:1:0): dev_active 0
> > (pass1:mrsas0:0:1:0): devq_openings 1024
> > (pass1:mrsas0:0:1:0): devq_queued 0
> > (pass1:mrsas0:0:1:0): held 0
> > (pass1:mrsas0:0:1:0): mintags 2
> > (pass1:mrsas0:0:1:0): maxtags 1024
> >
> > Value 1024 is hard coded for my testing. In MegaRaid controller and SAS-
> > HBA Driver read max commands value from FW.
> > Similar to "RequestCredit".. Different FW has different value, but they
> > are every time above 255.
> >
> >
> > When I run IOs dev_active stays in range of 0-255 only. See below
> > output when I run IOs on /dev/da1 and /dev/da13. I expect total
> > dev_openings should go beyond 255, which is not happening.
> >
> >
> > dhcp-135-24-192-127# camcontrol tags da1 -v
> > (pass1:mrsas0:0:1:0): dev_openings 832
> > (pass1:mrsas0:0:1:0): dev_active 192
> > (pass1:mrsas0:0:1:0): devq_openings 832
> > (pass1:mrsas0:0:1:0): devq_queued 0
> > (pass1:mrsas0:0:1:0): held 0
> > (pass1:mrsas0:0:1:0): mintags 2
> > (pass1:mrsas0:0:1:0): maxtags 1024
> > dhcp-135-24-192-127# camcontrol tags da13 -v
> > (pass13:mrsas0:0:13:0): dev_openings 881
> > (pass13:mrsas0:0:13:0): dev_active 143
> > (pass13:mrsas0:0:13:0): devq_openings 881
> > (pass13:mrsas0:0:13:0): devq_queued 0
> > (pass13:mrsas0:0:13:0): held 0
> > (pass13:mrsas0:0:13:0): mintags 2
> > (pass13:mrsas0:0:13:0): maxtags 1024
> >
> >
> >
> >
> > Jim:
> > Below is my API call. I have hard code value "queue_depth" = 1024
> >
> > sc->sim_0 = cam_sim_alloc(mrsas_action, mrsas_poll, "mrsas", sc,
> > device_get_unit(sc->mrsas_dev), &sc->sim_lock, queue_depth,
> > queue_depth, devq);
> >
> > ~ Kashyap
> >
> > >
> > > Ken
> > > --
> > > Kenneth Merry
> > > ken at FreeBSD.ORG
> > _______________________________________________
> > freebsd-scsi at freebsd.org mailing list
> > http://lists.freebsd.org/mailman/listinfo/freebsd-scsi
> > To unsubscribe, send any mail to "freebsd-scsi-unsubscribe at freebsd.org"
--
Kenneth Merry
ken at FreeBSD.ORG
More information about the freebsd-scsi
mailing list