> -----Original Message----- > From: [email protected] [mailto:[email protected]] > On Behalf Of Joe Hoot > Sent: Friday, April 29, 2011 11:47 AM > To: [email protected] > Subject: Re: determining window size > > We have a few environments that we're testing this with right now. We > are mostly using bnx2 (the nics are capable of offloading, but we > aren't using that portion). Since you can't disable the nic from doing > offloading from the BIOS, we do ti when the system comes only with the > following: > Not really.. You are not using the offload stack when you configure your iscsi sessions to use TCP in the OS. Once the BIOS passes control to the OS, the OS can create sessions through either stack. In fact even if you enabled BIOS(offload mode) you can't really offload iSCSI unless there is a LICENSE key plugged into the motherboard.
I presume that is the case unless you created session using the iface created by bnx2i.. > for ETH in eth0 eth1 eth2 eth3; do > ethtool -K $ETH tx off > ethtool -K $ETH rx off > ethtool -K $ETH sg off > ethtool -K $ETH tso off > ethtool -K $ETH gso off > done This is not necessary to offload infact you may have a performance hit since you do this.. > > For some reason, bnx2i is still getting loaded in the OS as a module, > but (I believe, because I tested this awhile ago) even if I rmmod > bnx2i, it will automatically get loaded again once iscsid starts back > up. Even this is not necessary.. Since even if it is loaded it is not getting used.. I guess.. the trace will exactly point to the tcp window sizes here... Look for resets either by initiator or target a zero window could be pretty near to a reset.. Sometimes if either initiator or target advertizes a zero window it results in TCP resets and therefore an error messages in the kernel. So watch for those kernel messages as well to profile where you need to search in the trace. > > > On Apr 29, 2011, at 10:46 AM, <[email protected]> > <[email protected]> wrote: > > > > > > >> -----Original Message----- > >> From: [email protected] [mailto:open- > [email protected]] > >> On Behalf Of Joe Hoot > >> Sent: Thursday, April 28, 2011 8:08 PM > >> To: [email protected] > >> Subject: Re: determining window size > >> > >> Thanks for the quick response (as always) Mike. > >> > >> Ok. I see where you're going. I should be able to build a kernel, > but > >> I have already veered a little bit away from where Oracle already is > >> with their OVM kernel. I don't want to veer to far (even with > printk > >> statements)-- but maybe... let me think about possibly doing > something > >> with just tcpdump to see if I can see tcp windows scaling real low. > >> > >> We have a concern at ITEC that the EqualLogic is setting TCP windows > >> towards zero. And when it does that we are concerned that if it > does > >> this too much, even noop pings aren't getting through. > > > > What is the NIC at the initiator..? Is it an offload session ? > > > > I have seen a TCP zero window issue recently. > > > >> > >> So... let think about this and see if I can respond again to you > >> tomorrow morning. > >> > >> Thanks again for jumping on this. > >> > >> Joe > >> > >> > >> On Apr 28, 2011, at 6:29 PM, Mike Christie wrote: > >> > >>> On 04/28/2011 02:42 PM, Joe Hoot wrote: > >>>> Mike or others, > >>>> > >>>> I would like to enable a small (yes, feel free to snicker here...) > >> bit of debugging on one of our systems to try to capture tcp window > >> sizes (so that I can attempt to trend some behaviors). Mike, recall > >> awhile ago you had mentioned that the following will help us gather > >> more information: > >>>> > >>> > >>> How are you setting the window sizes? > >>> > >>>> echo 1> /sys/module/libiscsi2/parameters/debug_libiscsi > >>>> echo 1> /sys/module/libiscsi_tcp/parameters/debug_libiscsi_tcp > >>>> echo 1> /sys/module/iscsi_tcp/parameters/debug_iscsi_tcp > >>>> > >>>> I believe that the only one that I really need to turn on to get > >> more information about size of data being sent would be this one: > >>>> > >>>> echo 1> /sys/module/libiscsi_tcp/parameters/debug_libiscsi_tcp > >>>> > >>>> as it gives me something like this when I run this command: > >>>> [root@oim61024006 ~]# echo 1> > >> /sys/module/libiscsi_tcp/parameters/debug_libiscsi_tcp; sleep 1; dd > >> if=/dev/zero > >> of=/var/ovs/mount/DCA44DBDD7944327A7945E96BF8F7CCC/joetest.dd.img > bs=1K > >> count=1; echo 0> > >> /sys/module/libiscsi_tcp/parameters/debug_libiscsi_tcp > >>>> 1+0 records in > >>>> 1+0 records out > >>>> 1024 bytes (1.0 kB) copied, 3.7e-05 seconds, 27.7 MB/s > >>>> [root@oim61024006 ~]# > >>>> > >>>> That produced this in /var/log/messages: > >>>> Apr 28 15:26:07 oim61024006 kernel: connection22:0: > >> iscsi_tcp_task_init task deq [itt 0x1c6 imm 0 unsol 0] > >>>> Apr 28 15:26:07 oim61024006 kernel: connection22:0: > >> iscsi_tcp_segment_done copied 0 0 size 48 xmit > >>>> Apr 28 15:26:07 oim61024006 kernel: connection22:0: > >> iscsi_tcp_segment_done copied 0 48 size 48 xmit > >>>> Apr 28 15:26:07 oim61024006 kernel: connection22:0: > >> iscsi_tcp_segment_done total copied 48 total size 48 > >>> > >>> Need libiscsi debugging on, but this looks like it might have been > a > >>> iscsi pdus related to scsi READ commands. On the next line we see > >> some > >>> data coming in from the target in response to some command. > >>> > >>> > >>>> Apr 28 15:26:07 oim61024006 kernel: connection22:0: > >> iscsi_tcp_recv_skb in 560 bytes > >>>> Apr 28 15:26:07 oim61024006 kernel: connection22:0: > >> iscsi_tcp_recv_skb skb deace200 ptr=dd84c800 avail=560 > >>>> Apr 28 15:26:07 oim61024006 kernel: connection22:0: > >> iscsi_tcp_segment_done copied 0 0 size 48 recv > >>>> Apr 28 15:26:07 oim61024006 kernel: connection22:0: > >> iscsi_tcp_segment_recv copying 48 > >>>> Apr 28 15:26:07 oim61024006 kernel: connection22:0: > >> iscsi_tcp_segment_done copied 0 48 size 48 recv > >>>> Apr 28 15:26:07 oim61024006 kernel: connection22:0: > >> iscsi_tcp_segment_done total copied 48 total size 48 > >>>> Apr 28 15:26:07 oim61024006 kernel: connection22:0: > >> iscsi_tcp_recv_skb segment done > >>>> > >>>> ---> and I'm thinking that this is actually where my dd starts > >> going: > >>>> Apr 28 15:26:07 oim61024006 kernel: connection22:0: > >> iscsi_tcp_hdr_dissect opcode 0x25 ahslen 0 datalen 512 > >>> > >>> opcode 25 is a iSCSI SCSI DATA_IN. Basically the target sends data > to > >>> the initiator in this type of pdu in response to a read command. > >> Below > >>> we see more data getting read in. > >>> > >>>> Apr 28 15:26:07 oim61024006 kernel: connection22:0: > >> iscsi_tcp_hdr_dissect iscsi_tcp_begin_data_in( offset=0, > datalen=512) > >>> > >>> so here datalen means that the DATA_IN pdu has 512 bytes of data > that > >> we > >>> are going to read in. > >>> > >>>> Apr 28 15:26:07 oim61024006 kernel: connection22:0: > >> iscsi_tcp_recv_skb in 512 bytes > >>>> Apr 28 15:26:07 oim61024006 kernel: connection22:0: > >> iscsi_tcp_recv_skb skb deace200 ptr=dd84c830 avail=512 > >>>> Apr 28 15:26:07 oim61024006 kernel: connection22:0: > >> iscsi_tcp_segment_done copied 0 0 size 512 recv > >>>> Apr 28 15:26:07 oim61024006 kernel: connection22:0: > >> iscsi_tcp_segment_recv copying 512 > >>>> Apr 28 15:26:07 oim61024006 kernel: connection22:0: > >> iscsi_tcp_segment_done copied 0 512 size 512 recv > >>>> Apr 28 15:26:07 oim61024006 kernel: connection22:0: > >> iscsi_tcp_segment_done total copied 512 total size 512 > >>>> Apr 28 15:26:07 oim61024006 kernel: connection22:0: > >> iscsi_tcp_recv_skb segment done > >>> > >>> So the above blob just means we got the 512 bytes read in > >> successfully. > >>> > >>>> Apr 28 15:26:07 oim61024006 kernel: connection22:0: > >> iscsi_tcp_hdr_recv_prep (digest disabled) > >>>> Apr 28 15:26:07 oim61024006 kernel: connection22:0: > >> iscsi_tcp_recv_skb in 0 bytes > >>>> Apr 28 15:26:07 oim61024006 kernel: connection22:0: > >> iscsi_tcp_recv_skb no more data avail. Consumed 0 > >>>> --> and this is where it had finished? > >>>> > >>>> Apr 28 15:26:07 oim61024006 kernel: connection22:0: > >> iscsi_tcp_segment_done copied 0 0 size 0 xmit > >>>> Apr 28 15:26:07 oim61024006 kernel: connection22:0: > >> iscsi_tcp_segment_done total copied 0 total size 0 > >>>> > >>>> I'm assuming that these "size 48 recv" might be some keepalive > thing > >> (nop pings, dm-multipath readsector 0, etc..)? > >>>> > >>> > >>> Is a iscsi header. In a iscsi packet there is basically a header > that > >>> tells us what type of command it is (ping, scsi command, ready to > >>> transfer, login, etc), how much data is going to be transferred in > >> the > >>> data segment after the header, etc. > >>> > >>> For the xmit path you need libiscsi debugging on as well as > >> libiscsi_tcp > >>> to determine what type of iscsi header it is. > >>> > >>> See above for info the packets. > >>> > >>>> 3) "iscsi_tcp_recv_skb in 512 bytes" > >>>> > >>>> skb = send kb's? > >>> > >>> skb is a networking thing. It is just short for "struct sk_buff" > >> which > >>> just represents a buffer that we could send/recv to/from. > >>> > >>> > >>>> > >>>> etc... > >>>> > >>>> But what I'm really trying to pull out of this information is > trying > >> to trend how often the initiator is trying to xmit data, but is not > >> allowed to. Maybe there is also a way (outside of just > tcpdumping... > >> or even with tcpdump, with a tweaked filter so I don't get gigs of > >> pcaps) to understand what the tcp window size is currently being set > >> to. > >>>> > >>> > >>> There is no nice printk for when sendpage/sendmsg has failed to > send > >>> data. In the past logs I could tell it was happening because in the > >> xmit > >>> messages we would repeat messages like this: > >>> > >>> connection7:0: iscsi_tcp_segment_done copied 2592 0 size 4096 xmit > >>> > >>> which comes from here: > >>> > >>> ISCSI_DBG_TCP(tcp_conn->iscsi_conn, "copied %u %u size %u > >> %s\n", > >>> segment->copied, copied, segment->size, > >>> recv ? "recv" : "xmit"); > >>> > >>> And so the first value is how much we have copied so far. The > second > >>> values is how much got copied when we just called sendpage/sendmsg > >> and > >>> the last value size is the total amount of data we want to send. > >>> > >>> In the logs you sent it kept repeating the same values and we never > >> saw > >>> the copied values get incremented and we never saw this message > >>> indicating we sent everything. > >>> > >>> /* Do we have more scatterlist entries? */ > >>> ISCSI_DBG_TCP(tcp_conn->iscsi_conn, "total copied %u total > >> size > >>> %u\n", > >>> segment->total_copied, segment->total_size); > >>> > >>> > >>>> I would then try to parse it after the fact to put the data in rrd > >> graphs. > >>>> > >>>> If there are any sites that explain this in more detail, I would > >> happy to dig. The only issue I have is that I'm not a programmer... > I > >> haven't reviewed or coded in C in about 10 years (and even then I > was > >> just learning it in school). So if the site is too in depth, I may > >> still be asking questions... but regardless, I am more than happy to > >> hit the ground running and at least do the due diligence of > reviewing > >> the info if someone can point me in the right direction. > >>>> > >>> > >>> If you can build I kernel I can make this a lot easier. I can just > >> add a > >>> printk to the sendpage/sendmsg call, and so it will be really easy > to > >>> pinpoint. > >>> > >>> -- > >>> You received this message because you are subscribed to the Google > >> Groups "open-iscsi" group. > >>> To post to this group, send email to [email protected]. > >>> To unsubscribe from this group, send email to open- > >> [email protected]. > >>> For more options, visit this group at > >> http://groups.google.com/group/open-iscsi?hl=en. > >>> > >> > >> -- > >> You received this message because you are subscribed to the Google > >> Groups "open-iscsi" group. > >> To post to this group, send email to [email protected]. > >> To unsubscribe from this group, send email to open- > >> [email protected]. > >> For more options, visit this group at > >> http://groups.google.com/group/open-iscsi?hl=en. > > > > -- > > You received this message because you are subscribed to the Google > Groups "open-iscsi" group. > > To post to this group, send email to [email protected]. > > To unsubscribe from this group, send email to open- > [email protected]. > > For more options, visit this group at > http://groups.google.com/group/open-iscsi?hl=en. > > > > -- > You received this message because you are subscribed to the Google > Groups "open-iscsi" group. > To post to this group, send email to [email protected]. > To unsubscribe from this group, send email to open- > [email protected]. > For more options, visit this group at > http://groups.google.com/group/open-iscsi?hl=en. -- You received this message because you are subscribed to the Google Groups "open-iscsi" group. To post to this group, send email to [email protected]. To unsubscribe from this group, send email to [email protected]. For more options, visit this group at http://groups.google.com/group/open-iscsi?hl=en.
