Re: [SPDK] BDEV-IO Lifecycle - Need your input.

[Harris, James R <james.r.harris at intel.com>]

[-- Attachment #1: Type: text/plain, Size: 12026 bytes --]

Thanks Srikanth.  Sounds like the spdk_bdev_io pool sizing along with spdk_bdev_queue_io_wait() meets your needs then regarding the spdk_bdev_io memory.

Regarding zcopy_start/zcopy_end – it looks like you’ve already added a bunch of comments to Ben’s patch on GerritHub.  For now I’d say let’s continue our discussion there.  I’ve responded to a couple of similar questions there and I’m sure Ben will have more replies tomorrow.

-Jim


From: "Kaligotla, Srikanth" <Srikanth.Kaligotla(a)netapp.com>
Date: Thursday, June 21, 2018 at 5:01 PM
To: James Harris <james.r.harris(a)intel.com>, "Walker, Benjamin" <benjamin.walker(a)intel.com>, Daniel Verkamp <daniel.verkamp(a)intel.com>
Cc: "raju.gottumukkala(a)broadcom.com" <raju.gottumukkala(a)broadcom.com>, "Meneghini, John" <John.Meneghini(a)netapp.com>, "Rodriguez, Edwin" <Ed.Rodriguez(a)netapp.com>, "Pai, Madhu" <Madhusudan.Pai(a)netapp.com>, "NGC-john.barnard-broadcom.com" <john.barnard(a)broadcom.com>, "spdk(a)lists.01.org" <spdk(a)lists.01.org>
Subject: RE: BDEV-IO Lifecycle - Need your input.

Hi Jim,

I wish I joined the community meeting, I also missed the spdk_bdev_queue_io_wait().

So, there are 2 issues I am intending to solve,

1.       We want to ensure that an instance of bdev-io is acquired prior to or along with I/O data buffer. This allows for better error handling when bdev_io pool is exhausted. Yes, we can solve it via sizing the pool right. The changes queue-io-wait also addresses the problem.
2.       Most importantly, by extending the life of bdev-io (Same life span as nvmf_request structure), abort use case and other use cases that involve cleaning up of IO data buffer are accomplished effectively. Let me elaborate; bdev-io is the fabric that connects the nvme command and operation with the backend. The driver context and data buffer context is stored in bdev-io. The freeing up of bdev_io resource is pushed to the end so the I/O cleanup can happen after controller has transmitted the data to the host. In absence of bdev_io, we will have end up adding more and more void context in request structure. Hence the push to extend the life of bdev_io.

I see Ben’s recent patch “zcopy_start” and “zcopy_end”; We can make that work as long as the bdev-io allocated/acquired stays till the end. One of the challenges I see with that patch is defining a callback for the I/O submission. For instance, zopy_start will allocate bdev_io and submits the I/O to the bdev device. Underlying implementation can be synchronous or can be asynchronous. The callback for this submission should check to see if it is a BUFFER-READY or PENDING and accordingly relay it back to transport. Next phase is actual I/O submission. Let’s say, it reuses the bdev-io obtained in ZCOPY-START, now the callback should determine if it is a success or failure. Finally when ZCOPY-END is invoked the supplied bdev_io will have all necessary data to release the WRITE buffer or unlock the read buffer based on the operation performed.

I hope I’m making sense. I guess, my effort is to extend the lifespan of bdev-io and let it continue to host the driver context and buffer context populated during the BEGIN phase.

Regards,
Srikanth

From: Harris, James R <james.r.harris(a)intel.com>
Sent: Thursday, June 21, 2018 2:21 PM
To: Kaligotla, Srikanth <Srikanth.Kaligotla(a)netapp.com>; Walker, Benjamin <benjamin.walker(a)intel.com>; Verkamp, Daniel <daniel.verkamp(a)intel.com>
Cc: raju.gottumukkala(a)broadcom.com; Meneghini, John <John.Meneghini(a)netapp.com>; Rodriguez, Edwin <Ed.Rodriguez(a)netapp.com>; Pai, Madhu <Madhusudan.Pai(a)netapp.com>; NGC-john.barnard-broadcom.com <john.barnard(a)broadcom.com>; spdk(a)lists.01.org
Subject: Re: BDEV-IO Lifecycle - Need your input.

Hi Srikanth,

Following up on this thread and the discussion in yesterday’s community meeting.

The recent spdk_bdev_queue_io_wait() changes allow an SPDK application to work better in general when the spdk_bdev_io buffer pool is exhausted.  We still need to make changes to the NVMe-oF target to use this new API but that should be done soon-ish.

With that in place, the spdk_bdev_io buffer pool itself can be configured up or down when the application starts.  Currently the default is 64K spdk_bdev_io buffers.  sizeof(struct spdk_bdev_io) == 216 plus the per-IO context size allocated for the bdev module.  This can be up to 192 bytes (virtio bdev module) but is likely much smaller for you depending on the context size for your ontap bdev module.

Let’s assume your per-IO context size is 64 bytes.  64K x (192 + 64) = 16MB.

I’m not sure how many spdk_bdev_io you need in flight at any given time.  64K seems like a lot but I’d be curious to hear your thoughts on this.  If this is the right number, then worst case, there would be about 16MB of DRAM that would sit unused if the NVMe-oF target in your system was not active.  Is that too burdensome for your application?

Thanks,

-Jim



From: "Kaligotla, Srikanth" <Srikanth.Kaligotla(a)netapp.com<mailto:Srikanth.Kaligotla(a)netapp.com>>
Date: Wednesday, June 20, 2018 at 12:47 PM
To: "Walker, Benjamin" <benjamin.walker(a)intel.com<mailto:benjamin.walker(a)intel.com>>, James Harris <james.r.harris(a)intel.com<mailto:james.r.harris(a)intel.com>>, Daniel Verkamp <daniel.verkamp(a)intel.com<mailto:daniel.verkamp(a)intel.com>>
Cc: "raju.gottumukkala(a)broadcom.com<mailto:raju.gottumukkala(a)broadcom.com>" <raju.gottumukkala(a)broadcom.com<mailto:raju.gottumukkala(a)broadcom.com>>, "Meneghini, John" <John.Meneghini(a)netapp.com<mailto:John.Meneghini(a)netapp.com>>, "Rodriguez, Edwin" <Ed.Rodriguez(a)netapp.com<mailto:Ed.Rodriguez(a)netapp.com>>, "Pai, Madhu" <Madhusudan.Pai(a)netapp.com<mailto:Madhusudan.Pai(a)netapp.com>>, "NGC-john.barnard-broadcom.com" <john.barnard(a)broadcom.com<mailto:john.barnard(a)broadcom.com>>, "spdk(a)lists.01.org<mailto:spdk(a)lists.01.org>" <spdk(a)lists.01.org<mailto:spdk(a)lists.01.org>>
Subject: RE: BDEV-IO Lifecycle - Need your input.

Hello,

First revision of changes to extend the lifecycle of bdev_io are available for review. I would like to solicit your input on the proposed API/Code flow changes.

https://review.gerrithub.io/c/spdk/spdk/+/415860

Thanks,
Srikanth


From: "Kaligotla, Srikanth" <Srikanth.Kaligotla(a)netapp.com<mailto:Srikanth.Kaligotla(a)netapp.com>>
Date: Friday, May 11, 2018 at 2:27 PM
To: "Walker, Benjamin" <benjamin.walker(a)intel.com<mailto:benjamin.walker(a)intel.com>>, "Harris, James R" <james.r.harris(a)intel.com<mailto:james.r.harris(a)intel.com>>
Cc: "raju.gottumukkala(a)broadcom.com<mailto:raju.gottumukkala(a)broadcom.com>" <raju.gottumukkala(a)broadcom.com<mailto:raju.gottumukkala(a)broadcom.com>>, "Meneghini, John" <John.Meneghini(a)netapp.com<mailto:John.Meneghini(a)netapp.com>>, "Rodriguez, Edwin" <Ed.Rodriguez(a)netapp.com<mailto:Ed.Rodriguez(a)netapp.com>>, "Pai, Madhu" <Madhusudan.Pai(a)netapp.com<mailto:Madhusudan.Pai(a)netapp.com>>, "NGC-john.barnard-broadcom.com" <john.barnard(a)broadcom.com<mailto:john.barnard(a)broadcom.com>>, "spdk(a)lists.01.org<mailto:spdk(a)lists.01.org>" <spdk(a)lists.01.org<mailto:spdk(a)lists.01.org>>
Subject: RE: BDEV-IO Lifecycle - Need your input.

CC: List

Hi Ben,

Your proposal<https://review.gerrithub.io/#/c/spdk/spdk/+/386166/> to interface with the backend to acquire and release buffers is good. You have accurately stated that the challenge is in developing intuitive semantics. And that has been my struggle. To me, there are two problem statements;

1.       It is expected that bdev_io pool is sized correctly so the call to get bdev_io succeeds. Failure to acquire bdev_io will result in DEVICE-ERROR. The transport is already capable of handling temporary memory failures by moving the request to PENDING-Queue. Hence the proposal to change the bdev_io lifecycle and perhaps connect it with the spdk_nvmf_request object. Thus all buffer needs are addressed at the beginning of I/O request.
2.       I/O Data buffers are sourced and managed by the backend. One of the challenges I see with your proposed interface is the lack of details like if resource is being acquired for a READ operation or WRITE operation. The handling is quite different in each case. Since bdev_io->type is overloaded the type of I/O operation is lost. I suppose one can cast the cb_arg (nvmf_request) and then proceed. Also, the bdev_io should be present to RELEASE the buffer. Zero copy semantics warrants that data buffer stays until controller-to-host has occurred. In other words, bdev_io lives till REQUEST come to COMPLETE state.

What are your thoughts in introducing spdk_bdev_init() and spdk_bdev_fini() as an alternative approach to extend the lifecyle of bdev_io and allow data buffer management via bdev fn_table ?

I hope I’m making sense…

Thanks,
Srikanth

From: Walker, Benjamin <benjamin.walker(a)intel.com<mailto:benjamin.walker(a)intel.com>>
Sent: Friday, May 11, 2018 12:28 PM
To: Harris, James R <james.r.harris(a)intel.com<mailto:james.r.harris(a)intel.com>>; Kaligotla, Srikanth <Srikanth.Kaligotla(a)netapp.com<mailto:Srikanth.Kaligotla(a)netapp.com>>
Cc: raju.gottumukkala(a)broadcom.com<mailto:raju.gottumukkala(a)broadcom.com>; Meneghini, John <John.Meneghini(a)netapp.com<mailto:John.Meneghini(a)netapp.com>>; Rodriguez, Edwin <Ed.Rodriguez(a)netapp.com<mailto:Ed.Rodriguez(a)netapp.com>>; Pai, Madhu <Madhusudan.Pai(a)netapp.com<mailto:Madhusudan.Pai(a)netapp.com>>; NGC-john.barnard-broadcom.com <john.barnard(a)broadcom.com<mailto:john.barnard(a)broadcom.com>>
Subject: Re: BDEV-IO Lifecycle - Need your input.

Hi Srikanth,

Yes - we'll need to introduce some way to acquire and release buffers from the bdev layer earlier on in the state machine that processes an NVMe-oF request. I've had this patch out for review for several months as a proposal for this scenario:

https://review.gerrithub.io/#/c/spdk/spdk/+/386166/

It doesn't pass the tests - it's just a proposal for the interface. 90% of the challenge here is in developing intuitive semantics.

Thanks,
Ben

P.S. This is the kind of discussion that would fit perfectly on the mailing list.

On Wed, 2018-05-09 at 20:35 +0000, Kaligotla, Srikanth wrote:
Hi Ben, Hi James,

I would like to solicit opinions on the lifecycle for bdev-io resource object. Attached is the image of RDMA state machine in its current implementation. When the REQUEST enters the NEED-BUFFER state, the buffers necessary for carrying out I/O operation are allocated/acquired from the memory pool. An instance of BDEV-IO comes into existence after REQUEST reaches READY-TO-EXECUTE state. The BDEV-IO is teared down as soon the backend returns. From a BDEV perspective, BDEV-IO is simply a translation unit that facilitates I/O buffers from the backend. The driver context embedded within bdev_io holds great deal of information pertaining to the I/O under execution. It assists in error handling, dereferencing the buffers upon I/O completion and in abort handling. In summary, the bdev_io stays alive until request has come to COMPLETE state. I’d like to hear peoples thoughts in introducing the plumbing to acquire BDEV-IO resource in REQUEST-NEED-BUFFER state and release it in REQUEST-COMPLETE state.  I will shortly have patch available for review that introduces spdk_bdev_init and spdk_bdev_fini which in turn invokes the corresponding bdev fn_table to  initialize/cleanup.

I wanted to use this email to communicate our intent and solicit your feedback. We have a working implementation of the above proposal and prior to pushing it upstream for review would like to hear your thoughts. These proposed changes to upstream are a result of FC transport work in collaboration with the Broadcom team who are also copied to this mail. Myself and John will be at SPDK Dev conference and if required we can elaborate further on this proposal.

Thanks,
Srikanth


[-- Attachment #2: attachment.html --]
[-- Type: text/html, Size: 28484 bytes --]