Date: Mon, 6 Jan 2014 10:29:35 -0800 From: Doug Ambrisko <ambrisko@cisco.com> To: "Desai, Kashyap" <Kashyap.Desai@lsi.com> Cc: "scottl@netflix.com" <scottl@netflix.com>, "Radford, Adam" <Adam.Radford@lsi.com>, "Kenneth D. Merry" <ken@kdm.org>, "sean_bruno@yahoo.com" <sean_bruno@yahoo.com>, "Mankani, Krishnaraddi" <Krishnaraddi.Mankani@lsi.com>, "dwhite@ixsystems.com" <dwhite@ixsystems.com>, "Maloy, Joe" <Joe.Maloy@lsi.com>, "jpaetzel@freebsd.org" <jpaetzel@freebsd.org>, "freebsd-scsi@freebsd.org" <freebsd-scsi@freebsd.org>, "McConnell, Stephen" <Stephen.McConnell@lsi.com> Subject: Re: LSI - MR-Fusion controller driver <mrsas> patch and man page Message-ID: <20140106182935.GC93278@cisco.com> In-Reply-To: <8c423414ecc2421fbace3eb9f386be91@BN1PR07MB247.namprd07.prod.outlook.com> References: <08ba2a262fba45f687cdd3225f325110@BN1PR07MB247.namprd07.prod.outlook.com> <20140103211449.GA69721@cisco.com> <8c423414ecc2421fbace3eb9f386be91@BN1PR07MB247.namprd07.prod.outlook.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Mon, Jan 06, 2014 at 07:43:27AM +0000, Desai, Kashyap wrote: | Doug: | | I have replied inline for your comment. For most of the comment you | posted below can be work out and we can prioritize completion of those | based on nature of issue. | I will work on READ/WRITE comparison w.r.t <mfi> and <mrsas> as a | priority item. IO performance is the highest priority. | Please consider that it will be really helpful for further development, | if we have common code check-in in Upstream FreeBSD. We will actively | work on each comment posted for <mrsas> and address those before we | have <mrsas> in-box for any FreeBSD release. What we have done in the past is to create a project branch to do this and it has worked well. It would be nice if you could check in directly to a this project branch so we could use svn to resolve differences. I poked around at adding the ioctl compatibility support and noticed that mrsas seems to lack a general method to send a dcmd to the controller. What Scott did in mfi was to create a generic function to send dcmd's with a simple data buffer. That data buffer was then put into the sg list in the generic function. This made it easier to translate the various ioctl's since we could convert the sg lists passed down from the user land into a simple malloced data base and then covert back out. In the mfituil ioctl method it just passes down the data buffer and then let's the driver convert it to a sg list. To make ioctl emulation easier in these cases: 32 bit FreeBSD LSI ioctl on 64 bit 32 bit Linux LSI ioctl on 32 bit 32 bit Linux LSI ioctl on 64 bit 32 bit mfiutil ioctl on 32 bit 32 bit mfiutil ioctl on 64 bit Of high priority is style, please make it style(9) compliant. There are a trailing spaces, spaces instead of tabs, I think lack of tab aligned indent. I noticed functions being externally defined in .c files instead of .h. There should be .h files that can be included in user land applications for ioctls and there can be .h files just for the kernel. These files will need to be installed as appropriate. I found this with the ioctl function. | > -----Original Message----- | > From: Doug Ambrisko [mailto:ambrisko@cisco.com] | > Sent: Saturday, January 04, 2014 2:45 AM | > To: Desai, Kashyap | > Cc: freebsd-scsi@freebsd.org; scottl@netflix.com; sean_bruno@yahoo.com; | > dwhite@ixsystems.com; jpaetzel@freebsd.org; Maloy, Joe; Mankani, | > Krishnaraddi; McConnell, Stephen; Saxena, Sumit; Radford, Adam; Kenneth | > D. Merry | > Subject: Re: LSI - MR-Fusion controller driver <mrsas> patch and man page | > | > On Fri, Dec 06, 2013 at 02:37:21PM +0000, Desai, Kashyap wrote: | > | Hi, | > | | > | Please consider attached patch for FreeBSD upstream check-in. | > | | > | Please find attached patch for <mrsas> driver for LSI's MegaRaid | > Controllers. This driver supports Thunderbolt onwards Device IDs of MR | > controllers. | > | Currently it supports 0x005B and 0x005D Device IDs. | > | | > | NOTE : This driver will not eliminate or by pass any functionality of <mfi> | > driver which already support above to Device IDs to keep existing user | > experience unchanged. | > | | > | <mfi> Driver will be always given priority over <mrsas> driver and | > | only if customer/user wants to use/migrate from <mfi> to <mrsas>, it | > | will hook up into kernel via device.hint rules. (Attached is mrsas man | > | page for more info.) | > | | > | LSI will continue to update <mrsas> driver in future in timely bases. | > | We have another set of patch in pipeline to be submitted for <mrsas>, | > | but need first go ahead for attached patch and later we will continue | > | to keep <mrsas> up-to-date (In sync with LSI released driver which is | > | available at lsi's external site) | > | | > | Apply patch with "patch -p0 < patchname.patch" from head directory. | > | | > | -- Few notes for user-- | > | LSI recommends using fusion_force bit In hint settings at start of the day, if | > they want to use <mrsas>. ( <mfi> will be a default choice for MR-Fusion | > HBA), if will be changed only with fusion_force hint settings. (See mrsas man | > page) Changing any default behavior is well tested for most of the condition. | > | Switching from <mfi> to <mrsas> for MR-Fusion options is designed to | > allow user as one time choice, though multiple time switch from <mfi> to | > <mrsas> is possible, it is not recommended. So, user needs to decide from | > start of the day, which driver they want to use for MR-Fusion card. | > | | > | -- Implementation details -- | > | To support this feature, we have modify <mfi> code to change default | > return type from probe. Currently <mfi> driver return | > "BUS_PROBE_DEFAULT". <mfi> driver has been be changed to return | > "BUS_PROBE_LOW_PRIORITY" if fusion_force hint from device.hints is set. | > | Please notice, above mentioned implementation in <mfi> driver is only | > applicable in case of MR-Fusion controller detection. For any other | > controllers, supported by <mfi> driver, the behavior of probe return will | > remain same as before. | > | | > | | > | -- High level feature list of <mrsas> -- 1. Supports Fast Path feature | > | of LSI controllers. | > | 2. Supports 4K sector Drives. | > | 3. CAM layer based interface. All VDs will be attached to CAM layer | > | (Expected storage will be visible in "camcontroll") 4. Complete | > | support of Online Controller Reset. (OCR) 5. OCR on Fimrware fault and IO | > timeout case. | > | 6. Work well with <storcli> management application which is generic | > application provided by LSI for all other Operating system. | > | 7. Supporst DIF enabled VDs (Same support as provided in Linux and | > | other OSes in FreeBSD) 8. Fast Path Load balance support. | > | | > | - In summary, this driver is in part with Linux based MR drivers and | > | all other features will be available to <mrsas> as planned activity | > | from LSI | > | | > | This code is well tested by LSI Q/A team on 32 bit and 64 bit FreeBSD | > Released OSes. | > | > Sorry it is has taken so long to start playing with this. I found one critical issue | > with read performance. I added a printf to the driver to make sure the fast | > path is not being used and it claims it isn't. | > Doing either a dd or using the | > dt(http://www.scsifaq.org/RMiller_Tools/dt.html) | > I see good write performance with both mrsas and mfi but on read mrsas is | > backing up and appears to be single threaded. With mrsas doing a dd | > of=/dev/zero if=/dev/da0 bs=1m I'm seeing 45685 kBps and with mfi seeing | > 597072 kBps via gstat. Using dt via: | > dt of=/dev/da0 bs=64k limit=1g aios=32 | > I see writes at 538977 Kbytes/sec and reads at bouncing around between | > 20000 - 8000 Kbytes/sec, mfi reports 244744 Kbytes/sec. The interesting | > piece is seeing the L(q) via gstat being around 30 when doing mrsas reads, | > writes are 1 or less. With mfi both read and writes are 1 or less. | > | Thanks for for providing feedback. This is very important observation. | We will work on this and get back to you. Meanwhile, if we have high | level acceptance(considering few defect fixes and optimization as a | follow up activity) of current <mrsas> code, can have <mrsas> commit in | current FreeBSD driver. This will definitely help us to manage source | code from single destination. Currently there are multiple flavors/version | of <mrsas> driver code is floated due to Upstream commit is pending. We | can work for each observation/optimization proposed from upstream | community on priority bases. Yes a project branch can help with this. | > This is with a GENERIC kernel on -current/amd64. I updated my source to | > 260231 and see the same. Witness, etc. are all on. I don't recall | > seeing this | > when I tested mrsas a long time ago. | | Thanks for additional info. | | > | > The card I'm using is: | > LSI MegaRAID SAS 9266-8i | > Firmware BIOS 5.38.00_4.12.05.00_0x05180000 | > Firmware BCON 6.1-49-e_49-Rel | > Firmware PCLI 05.05-03:#%00011 | > Firmware APP 3.220.75-2196 | > Firmware NVDT 2.1209.03-0117 | > Firmware BTBL 2.05.00.00-0010 | > Firmware BOOT 07.26.13.219 | > | > Some other issues I ran into when I kldunload mrsas I got a: | > sysctl_unregister_oid: failed to unregister sysctl and then when I | | I do not recall if we have seen this issue, but will try this on our setup. It think this and the panic might have been an artifact of my inition testing. I followed the instruction of kldload mrsas from /boot/loader.conf and saw this. However, when I started working the emulation I modified my GERNERIC config to not have mfi in it so I could switch between mfi and mrsas via kldload/kldunload to test ioctl emulation. When I did this I noticed that the patch included mrsas in the GENERIC configuration. That might have caused this issue having both the module and static driver in the same system. So this is probably a bad bug report. Doing kldload and kldunload without mrsas and mfi static in the kernel I didn't run into this again. | > kldload mrsas I got a panic: | > mrsas0: <LSI Thunderbolt SAS Controller> port 0x7000-0x70ff mem | > 0xdf060000-0xdf0 63fff,0xdf000000-0xdf03ffff irq 32 at device 0.0 on pci4 | > ######################################################### | > <mrsas> Driver has detected MR-Fusion Controller.If you wish to switch to | > <mfi> driver for MR-Fusion,read manual of mrsas. Once you chose to use | > <mrsas> Driverf or MR-Fusion Controllers, you should not switch to <mfi>.LSI | > recommend to use <m | > rsas> driver for MR-Fusion.Use <mrsas> if you are not sure about | > rsas> choice.Quick he | > lp to use <mfi> for MR-Fustion Card: | > Remove hint.mrsas.0.fusion_force=1 from /boot/device.hints. | > ######################################################### | > mrsas0: Waiting for FW to come to ready state | > | > <panic message from malloc> | > | > db> tr | > Tracing pid 534 tid 100245 td 0xfffff80015035490 | > malloc() at malloc+0x1c8/frame 0xfffffe03fbab3200 | > cpuctl_devs() at cpuctl_devs+0x61df/frame 0xfffffe03fbab3290 | > cpuctl_devs() at cpuctl_devs+0x6189/frame 0xfffffe03fbab32f0 | > cpuctl_devs() at cpuctl_devs+0x990a/frame 0xfffffe03fbab34b0 | > device_attach() at device_attach+0x3a2/frame 0xfffffe03fbab3510 | > pci_driver_added() at pci_driver_added+0xfa/frame 0xfffffe03fbab3550 | > devclass_driver_added() at devclass_driver_added+0x7d/frame | > 0xfffffe03fbab3590 | > devclass_add_driver() at devclass_add_driver+0x127/frame | > 0xfffffe03fbab35d0 | > module_register_init() at module_register_init+0xb0/frame | > 0xfffffe03fbab3600 | > linker_load_module() at linker_load_module+0xbda/frame | > 0xfffffe03fbab3930 | > kern_kldload() at kern_kldload+0xad/frame 0xfffffe03fbab3970 | > sys_kldload() at sys_kldload+0x5b/frame 0xfffffe03fbab39a0 | > amd64_syscall() at amd64_syscall+0x265/frame 0xfffffe03fbab3ab0 | > Xfast_syscall() at Xfast_syscall+0xfb/frame 0xfffffe03fbab3ab0 | > | > I played with storcli/storcli64 and MegaCli/MegaCli64 from LSI's web site. | > I noticed that the 32 bit versions are built for FreeBSD 8.1 and the 64 bit | > FreeBSD 7.4. It might not be a critical problem but some companies might still | > be on earlier releases. | > | > I need to look at adding the ioctl translation support for mfiutil and | > Linux to mrsas. This means we should have a 32 bit to 64 bit ioctl | > translator in the | > driver since our Linux emulation is 32 bit. | We can pick this work as next action item. I started working on this and mentioned some things that would make this easier. | > It's also handy to run 32 bit tools on the 64 bit kernel. | > | > It would be good to add alias support for /dev/mfid* such as was done via | > ATA CAM sys/cam/ata/ata_da.c. Search for ada_legacy_aliases and | > legacy_id. The helped the migration from /dev/ad* to /dev/ada*. | > | Let me check this. You can help me to understand what is expected here | and we can provide solution for this. | Again, we can mark this as next to-do for mrsas and will provide patch | as an when we have code ready. The idea is that when /dev/da<x> node is created and alias of /dev/mfi<y> is created point to /dev/da<x>. This way existing fstab that reference /dev/mfid0 for example will continue to work. It just makes it easier for migration. A trickier part will be the feature to allow both LD and PD of an LD to be available in a system. It is a feature that people use. With mfi it is easy to see RAID controlled devices via /dev/mfid<x> and mfisyspd<x>. The physical device access come up as /dev/da<x>. So people should know be very carefull when accessing /dev/da<x> since it could upset the RAID controller, such as flashing firmware and when the disk goes offline it can kick out the disk from the RAID. I haven't looked at how mrsas deals with deleting and recreating RAID on the fly. mfi can do it with the proper sysctl/tunables set. | > I'm not sure hint.mrsas.<x>.fusion_force="1" is a good toggle to | > enable/disable mrsas. I can see why you did that but it might be more | > obvious to make hint.mfi.<x>.disabled="1" to work since that follows more | > of what people are used to. | | Either way is fine, since hint.mfi.<X>.disabled may miss lead that we are | going to disable <mfi> completely. | We may have <mfi> driver working for Falcon and other old MFI based | controllers. Since it is a hint.mfi.<x>.disabled, the x refers to card x, ie. hint.mfi.1.disabled would disable the 2nd card in the system what would be the ThunderBolt based can and leave the others to be mfi. We should get more input from FreeBSD folks. fusion_force seems pretty non-obvious to me. I'd suggest force_mrsas to be more obvious and it should probably be a generic tunabled and device instance specific. Adding printing events to the kernel messages seems to be missing. This needs to be controlled by sysctl/tunable but I might not have seen this yet. I'd say the highest priority is on performance and style. Getting a review from Scott would be good but I'm not sure he has had time to do that. It might be good to look at mfi to see how things are laid out. FYI, my biggest problem working on FreeBSD is time. My day job priorities end up eating into time to work on "less critical" issues. Unfortunately a bunch of FreeBSD things fall into that catagory. I need to really resume on my mount path changes. I know what I need to do, I just need to carve out some time to work on it. Thanks, Doug A.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20140106182935.GC93278>