| Message ID | 20200923182021.3724-1-dmitry.fomichev@wdc.com |
|---|---|
| Headers | show |
| Series | hw/block/nvme: Support Namespace Types and Zoned Namespace Command Set | expand |
On Sep 24 03:20, Dmitry Fomichev wrote: > From: Niklas Cassel <niklas.cassel@wdc.com> > > In NVMe, a namespace is active if it exists and is attached to the > controller. > > CAP.CSS (together with the I/O Command Set data structure) defines what > command sets are supported by the controller. > > CC.CSS (together with Set Profile) can be set to enable a subset of the > available command sets. The namespaces belonging to a disabled command set > will not be able to attach to the controller, and will thus be inactive. > > E.g., if the user sets CC.CSS to Admin Only, NVM namespaces should be > marked as inactive. > Hmm. I'm not convinced that this is correct. Can you reference the spec? On the specific case you mention the spec is actually pretty clear: "When only the Admin Command Set is supported, any command submitted on an I/O Submission Queue and any I/O Command Set Specific Admin command submitted on the Admin Submission Queue is completed with status Invalid Command Opcode." My /interpretation/ (because the spec is vague on this point) is that with TP 4056, if the host writes 0x0 to CC.CSS, you will (should) just see Invalid Command Opcode for namespaces not supporting the NVM command set since we are operating in a backward compatible way. Now, if the host sets CC.CSS to 0x6, then it is obviously aware of namespaces and other rules apply. For instance, it may set the I/O Command Set Combination Index through a Set Features command, but TP 4056 is clear that the host will not be allowed to choose a combination that leaves an attached namespace unsupported. For this device, that does not implement namespace management and thus has no notion of attaching/detaching namespaces, the controller should by default choose an I/O Command Set Combination that indicates support for all I/O command sets that are required to support the namespaces configured.
On Thu, Sep 24, 2020 at 02:12:03PM +0200, Klaus Jensen wrote: > On Sep 24 03:20, Dmitry Fomichev wrote: > > From: Niklas Cassel <niklas.cassel@wdc.com> > > > > E.g., if the user sets CC.CSS to Admin Only, NVM namespaces should be > > marked as inactive. > > > > Hmm. I'm not convinced that this is correct. Can you reference the spec? > "If the user sets CC.CSS to Admin Only, NVM namespaces should be marked as inactive." After reading the spec several times, I agree that this statement is false, although I really wish it wasn't :) My interpretation was based on: From CC.CSS: "If bit 44 is set to ‘1’ in the Command Sets Supported (CSS) field, then the value 111b indicates that only the Admin Command Set is supported and that no I/O Command Set or I/O Command Set Specific Admin commands are supported." The NVM Command Set is a Command Set, so I assumed that since the Command Set was not supported, trying to do something like CNS 00h (Identify Namespace), would return an zero-filled struct. (Since the namespace belongs to a Command Set that is not supported.) To me it seems silly that a namespace can be "active" at the same time as the Command Set that the namespace belongs to is not enabled, but that seems like how the spec is written... Additionally in TP4056 section 5.19 it says: "If an attempt is made to attach a namespace to a controller that supports the corresponding I/O Command Set and the corresponding I/O Command Set is not enabled by the I/O Command Set profile feature, then the command shall be aborted with a status of I/O Command Set Not Enabled." But if you already have a e.g. a Key Value namespace attached, and boot up with e.g. CC.CSS = NVM, or CC.CSS = Admin, you will have a namespace belonging to a disabled command set attached (and "active" :p). But if you have no namespaces attached, boot up with CC.CSS = NVM or CC.CSS = Admin, you will not be allowed to attach your Key Value namespace. (5.19 says nothing about CC.CSS, but let's assume it isn't supposed to be allowed.) Will you be allowed to attach a ZNS namespace? (Since CC.CSS != ALL Selected, I would assume that is is not supposed to be allowed.) Now, will you be allowed to attach a NVM namespace when CC.CSS = Admin? (I assume that the intention is that you should.) CC.CSS can only be changed when the controller is disabled. I assume that you can attach NVM namespaces even when CC.CSS = Admin. (Attaching namespaces != NVM is probably not allowed when CC.CSS != ALL Selected.) Attaching namespaces are persistent. Even if you restart with CC.CSS = Admin, they will still be attached, and active. (Although you might not be able to do anything with them... see the next section.) You can not change to a profile (using Set Profile) that lacks a command set that you are currently using. This change is while the controller is enabled, so I guess it is ok that this check is stricter than CC.CSS. Things would have been way simpler if the controller just deattached non-supported namespaces internally at power on... > My /interpretation/ (because the spec is vague on this point) is that > with TP 4056, if the host writes 0x0 to CC.CSS, you will (should) just > see Invalid Command Opcode for namespaces not supporting the NVM command > set since we are operating in a backward compatible way. If a controller is booted with CC.CSS = NVM Command Set: We have an attached Key Value namespace. (It will be set as active.) I think that we can agree that any IOCS Key Value specific command will fail. We have an attached Zoned Namespace. (It will be set as active.) Sure, any IOCS Zoned Namespace Command specific command will fail. Your interpretation is that it will allow NVM Commands, since a Zoned Namespace implements the opcodes of the NVM Command Set. Each Command Set has a table with the opcodes that it supports. In e.g. Key Value Command Set, opcode 01h means "Store command" but in NVM Command Set, opcode 01h means "Write command". These commands are totally different, and uses completely different dwords. So I don't think that it is obvious that CC.CSS = NVM, should allow "NVM reserved opcodes" for certain Command Sets, but not in others. Like so many other things in NVMe, this will probably be vendor specific. For devices only supporting NVM + ZNS, it will probably support NVM Command Set commands even when CC.CSS = NVM, but it's probably not something that can be guarantee to be true for all devices supporting a Command Set that extends the NVM Command Set. Kind regards, Niklas
On Sep 24 18:17, Niklas Cassel wrote: > On Thu, Sep 24, 2020 at 02:12:03PM +0200, Klaus Jensen wrote: > > On Sep 24 03:20, Dmitry Fomichev wrote: > > > From: Niklas Cassel <niklas.cassel@wdc.com> > > > > > > E.g., if the user sets CC.CSS to Admin Only, NVM namespaces should be > > > marked as inactive. > > > > > > > Hmm. I'm not convinced that this is correct. Can you reference the spec? > > > > CC.CSS can only be changed when the controller is disabled. Right. I think I see you point. While the controller is disabled, the host obviously cannot even see what namespaces are available, so the controller is free to only expose (aka, attach) the namespaces that makes sense for the value of CC.CSS. OK then, the patch is good :)
On Thu, Sep 24, 2020 at 08:55:24PM +0200, Klaus Jensen wrote: > On Sep 24 18:17, Niklas Cassel wrote: > > On Thu, Sep 24, 2020 at 02:12:03PM +0200, Klaus Jensen wrote: > > > On Sep 24 03:20, Dmitry Fomichev wrote: > > > > From: Niklas Cassel <niklas.cassel@wdc.com> > > > > > > > > E.g., if the user sets CC.CSS to Admin Only, NVM namespaces should be > > > > marked as inactive. > > > > > > > > > > Hmm. I'm not convinced that this is correct. Can you reference the spec? > > > > > > > CC.CSS can only be changed when the controller is disabled. > > Right. I think I see you point. While the controller is disabled, the > host obviously cannot even see what namespaces are available, so the > controller is free to only expose (aka, attach) the namespaces that > makes sense for the value of CC.CSS. > > OK then, the patch is good :) That was my thought, that the controller internally would detach unsupported namespaces (even if the controller didn't expose namespace management capabilities to the user). This was how I assumed that things worked, but if we should follow the spec strictly, we should do like you suggested and keep them attached, and return the proper error code, on non-admin commands. Thank you for improving my understanding. Considering that CC.CSS can only be changed when the controller is disabled, I still kind of wished that the spec said that unsupported namespaces would be automatically detached. Kind regards, Niklas
On Sep 24 03:20, Dmitry Fomichev wrote: > v3 -> v4 > > - Fix bugs introduced in v2/v3 for QD > 1 operation. Now, all writes > to a zone happen at the new write pointer variable, zone->w_ptr, > that is advanced right after submitting the backend i/o. The existing > zone->d.wp variable is updated upon the successful write completion > and it is used for zone reporting. Some code has been split from > nvme_finalize_zoned_write() function to a new function, > nvme_advance_zone_wp(). > Same approach that I've used, +1. > - Make the code compile under mingw. Switch to using QEMU API for > mmap/msync, i.e. memory_region...(). Since mmap is not available in > mingw (even though there is mman-win32 library available on Github), > conditional compilation is added around these calls to avoid > undefined symbols under mingw. A better fix would be to add stub > functions to softmmu/memory.c for the case when CONFIG_POSIX is not > defined, but such change is beyond the scope of this patchset and it > can be made in a separate patch. > Ewwww. This feels like a hack or at the very least an abuse of the physical memory management API. If it really needs to be memory mapped, then I think a hostmem-based approach similar to what Andrzej did for PMR is needed (I think that will get rid of the CONFIG_POSIX ifdef at least, but still leave it slightly tricky to get it to work on all platforms AFAIK). But really, since we do not require memory semantics for this, then I think the abstraction is fundamentally wrong. I am, of course, blowing my own horn, since my implementation uses a portable blockdev for this. Another issue is the complete lack of endian conversions. Does it matter? It depends. Will anyone ever use this on a big endian host and move the meta data backing file to a little endian host? Probably not. So does it really matter? Probably not, but it is cutting corners. > - Make the list of review comments addressed in v2 of the series > (see below). > Very detailed! Thanks!
> -----Original Message----- > From: Klaus Jensen <its@irrelevant.dk> > Sent: Thursday, September 24, 2020 5:08 PM > To: Dmitry Fomichev <Dmitry.Fomichev@wdc.com> > Cc: Keith Busch <kbusch@kernel.org>; Klaus Jensen > <k.jensen@samsung.com>; Kevin Wolf <kwolf@redhat.com>; Philippe > Mathieu-Daudé <philmd@redhat.com>; Maxim Levitsky > <mlevitsk@redhat.com>; Fam Zheng <fam@euphon.net>; Niklas Cassel > <Niklas.Cassel@wdc.com>; Damien Le Moal <Damien.LeMoal@wdc.com>; > qemu-block@nongnu.org; qemu-devel@nongnu.org; Alistair Francis > <Alistair.Francis@wdc.com>; Matias Bjorling <Matias.Bjorling@wdc.com> > Subject: Re: [PATCH v4 00/14] hw/block/nvme: Support Namespace Types > and Zoned Namespace Command Set > > On Sep 24 03:20, Dmitry Fomichev wrote: > > v3 -> v4 > > > > - Fix bugs introduced in v2/v3 for QD > 1 operation. Now, all writes > > to a zone happen at the new write pointer variable, zone->w_ptr, > > that is advanced right after submitting the backend i/o. The existing > > zone->d.wp variable is updated upon the successful write completion > > and it is used for zone reporting. Some code has been split from > > nvme_finalize_zoned_write() function to a new function, > > nvme_advance_zone_wp(). > > > > Same approach that I've used, +1. > > > - Make the code compile under mingw. Switch to using QEMU API for > > mmap/msync, i.e. memory_region...(). Since mmap is not available in > > mingw (even though there is mman-win32 library available on Github), > > conditional compilation is added around these calls to avoid > > undefined symbols under mingw. A better fix would be to add stub > > functions to softmmu/memory.c for the case when CONFIG_POSIX is not > > defined, but such change is beyond the scope of this patchset and it > > can be made in a separate patch. > > > > Ewwww. > > This feels like a hack or at the very least an abuse of the physical > memory management API. > > If it really needs to be memory mapped, then I think a hostmem-based > approach similar to what Andrzej did for PMR is needed (I think that > will get rid of the CONFIG_POSIX ifdef at least, but still leave it > slightly tricky to get it to work on all platforms AFAIK). Ok, it looks that using the HostMemoryBackendFile backend will be more appropriate. This will remove the need for conditional compile. The mmap() portability is pretty decent across software platforms. Any poor Windows user who is forced to emulate ZNS on mingw will be able to do so, just without having zone state persistency. Considering how specialized this stuff is in first place, I estimate the number of users affected by this "limitation" to be exactly zero. > But really, > since we do not require memory semantics for this, then I think the > abstraction is fundamentally wrong. > Seriously, what is wrong with using mmap :) ? It is used successfully for similar applications, for example - https://github.com/open-iscsi/tcmu-runner/blob/master/file_zbc.c > I am, of course, blowing my own horn, since my implementation uses a > portable blockdev for this. > You are making it sound like the entire WDC series relies on this approach. Actually, the persistency is introduced in the second to last patch in the series and it only adds a couple of lines of code in the i/o path to mark zones dirty. This is possible because of using mmap() and I find the way it is done to be quite elegant, not ugly :) > Another issue is the complete lack of endian conversions. Does it > matter? It depends. Will anyone ever use this on a big endian host and > move the meta data backing file to a little endian host? Probably not. > So does it really matter? Probably not, but it is cutting corners. > Great point on endianness! Naturally, all file backed values are stored in their native endianness. This way, there is no extra overhead on big endian hardware architectures. Portability concerns can be easily addressed by storing metadata endianness as a byte flag in its header. Then, during initialization, the metadata validation code can detect the possible discrepancy in endianness and automatically convert the metadata to the endianness of the host. This part is out of scope of this series, but I would be able to contribute such a solution as an enhancement in the future. > > - Make the list of review comments addressed in v2 of the series > > (see below). > > > > Very detailed! Thanks!
On Sep 28 02:33, Dmitry Fomichev wrote: > > -----Original Message----- > > From: Klaus Jensen <its@irrelevant.dk> > > > > If it really needs to be memory mapped, then I think a hostmem-based > > approach similar to what Andrzej did for PMR is needed (I think that > > will get rid of the CONFIG_POSIX ifdef at least, but still leave it > > slightly tricky to get it to work on all platforms AFAIK). > > Ok, it looks that using the HostMemoryBackendFile backend will be > more appropriate. This will remove the need for conditional compile. > > The mmap() portability is pretty decent across software platforms. > Any poor Windows user who is forced to emulate ZNS on mingw will be > able to do so, just without having zone state persistency. Considering > how specialized this stuff is in first place, I estimate the number of users > affected by this "limitation" to be exactly zero. > QEMU is a cross platform project - we should strive for portability. Alienating developers that use a Windows platform and calling them out as "poor" is not exactly good for the zoned ecosystem. > > But really, > > since we do not require memory semantics for this, then I think the > > abstraction is fundamentally wrong. > > > > Seriously, what is wrong with using mmap :) ? It is used successfully for > similar applications, for example - > https://github.com/open-iscsi/tcmu-runner/blob/master/file_zbc.c > There is nothing fundamentally wrong with mmap. I just think it is the wrong abstraction here (and it limits portability for no good reason). For PMR there is a good reason - it requires memory semantics. > > I am, of course, blowing my own horn, since my implementation uses a > > portable blockdev for this. > > > > You are making it sound like the entire WDC series relies on this approach. > Actually, the persistency is introduced in the second to last patch in the > series and it only adds a couple of lines of code in the i/o path to mark > zones dirty. This is possible because of using mmap() and I find the way > it is done to be quite elegant, not ugly :) > No, I understand that your implementation works fine without persistance, but persistance is key. That is why my series adds it in the first patch. Without persistence it is just a toy. And the QEMU device is not just an "NVMe-version" of null_blk. And I don't think I ever called the use of mmap ugly. I called out the physical memory API shenanigans as a hack. > > Another issue is the complete lack of endian conversions. Does it > > matter? It depends. Will anyone ever use this on a big endian host and > > move the meta data backing file to a little endian host? Probably not. > > So does it really matter? Probably not, but it is cutting corners. > > After I had replied this, I considered a follow-up, because there are probably QEMU developers that would call me out on this. This definitely DOES matter to QEMU. > > Great point on endianness! Naturally, all file backed values are stored in > their native endianness. This way, there is no extra overhead on big endian > hardware architectures. Portability concerns can be easily addressed by > storing metadata endianness as a byte flag in its header. Then, during > initialization, the metadata validation code can detect the possible > discrepancy in endianness and automatically convert the metadata to the > endianness of the host. This part is out of scope of this series, but I would > be able to contribute such a solution as an enhancement in the future. > It is not out of scope. I don't see why we should merge something that is arguably buggy. Bottomline is that I just don't see why we should accept an implementation that a) excludes some platforms (Windows) from using persistence; and b) contains endianness conversion issues when there is a portable implementation posted that at least tries to convert endianness as needed.
On Mon, Sep 28, 2020 at 08:36:48AM +0200, Klaus Jensen wrote: > On Sep 28 02:33, Dmitry Fomichev wrote: > > You are making it sound like the entire WDC series relies on this approach. > > Actually, the persistency is introduced in the second to last patch in the > > series and it only adds a couple of lines of code in the i/o path to mark > > zones dirty. This is possible because of using mmap() and I find the way > > it is done to be quite elegant, not ugly :) > > > > No, I understand that your implementation works fine without > persistance, but persistance is key. That is why my series adds it in > the first patch. Without persistence it is just a toy. And the QEMU > device is not just an "NVMe-version" of null_blk. I really think we should be a bit more cautious of commiting to an on-disk format for the persistent state. Both this and Klaus' persistent state feels a bit ad-hoc, and with all the other knobs provided, it looks too easy to have out-of-sync states, or just not being able to boot at all if a qemu versions have different on-disk formats. Is anyone really considering zone emulation for production level stuff anyway? I can't imagine a real scenario where you'd want put yourself through that: you are just giving yourself all the downsides of a zoned block device and none of the benefits. AFAIK, this is provided as a development vehicle, closer to a "toy". I think we should consider trimming this down to a more minimal set that we *do* agree on and commit for inclusion ASAP. We can iterate all the bells & whistles and flush out the meta data's data marshalling scheme for persistence later.
On 2020/09/29 6:25, Keith Busch wrote: > On Mon, Sep 28, 2020 at 08:36:48AM +0200, Klaus Jensen wrote: >> On Sep 28 02:33, Dmitry Fomichev wrote: >>> You are making it sound like the entire WDC series relies on this approach. >>> Actually, the persistency is introduced in the second to last patch in the >>> series and it only adds a couple of lines of code in the i/o path to mark >>> zones dirty. This is possible because of using mmap() and I find the way >>> it is done to be quite elegant, not ugly :) >>> >> >> No, I understand that your implementation works fine without >> persistance, but persistance is key. That is why my series adds it in >> the first patch. Without persistence it is just a toy. And the QEMU >> device is not just an "NVMe-version" of null_blk. > > I really think we should be a bit more cautious of commiting to an > on-disk format for the persistent state. Both this and Klaus' persistent > state feels a bit ad-hoc, and with all the other knobs provided, it > looks too easy to have out-of-sync states, or just not being able to > boot at all if a qemu versions have different on-disk formats. > > Is anyone really considering zone emulation for production level stuff > anyway? I can't imagine a real scenario where you'd want put yourself > through that: you are just giving yourself all the downsides of a zoned > block device and none of the benefits. AFAIK, this is provided as a > development vehicle, closer to a "toy". > > I think we should consider trimming this down to a more minimal set that > we *do* agree on and commit for inclusion ASAP. We can iterate all the > bells & whistles and flush out the meta data's data marshalling scheme > for persistence later. +1 on this. Removing the persistence also removes the debate on endianess. With that out of the way, it should be straightforward to get agreement on a series that can be merged quickly to get developers started with testing ZNS software with QEMU. That is the most important goal here. 5.9 is around the corner, we need something for people to get started with ZNS quickly.
On Sep 28 22:54, Damien Le Moal wrote: > On 2020/09/29 6:25, Keith Busch wrote: > > On Mon, Sep 28, 2020 at 08:36:48AM +0200, Klaus Jensen wrote: > >> On Sep 28 02:33, Dmitry Fomichev wrote: > >>> You are making it sound like the entire WDC series relies on this approach. > >>> Actually, the persistency is introduced in the second to last patch in the > >>> series and it only adds a couple of lines of code in the i/o path to mark > >>> zones dirty. This is possible because of using mmap() and I find the way > >>> it is done to be quite elegant, not ugly :) > >>> > >> > >> No, I understand that your implementation works fine without > >> persistance, but persistance is key. That is why my series adds it in > >> the first patch. Without persistence it is just a toy. And the QEMU > >> device is not just an "NVMe-version" of null_blk. > > > > I really think we should be a bit more cautious of commiting to an > > on-disk format for the persistent state. Both this and Klaus' persistent > > state feels a bit ad-hoc, and with all the other knobs provided, it > > looks too easy to have out-of-sync states, or just not being able to > > boot at all if a qemu versions have different on-disk formats. > > > > Is anyone really considering zone emulation for production level stuff > > anyway? I can't imagine a real scenario where you'd want put yourself > > through that: you are just giving yourself all the downsides of a zoned > > block device and none of the benefits. AFAIK, this is provided as a > > development vehicle, closer to a "toy". > > > > I think we should consider trimming this down to a more minimal set that > > we *do* agree on and commit for inclusion ASAP. We can iterate all the > > bells & whistles and flush out the meta data's data marshalling scheme > > for persistence later. > > +1 on this. Removing the persistence also removes the debate on endianess. With > that out of the way, it should be straightforward to get agreement on a series > that can be merged quickly to get developers started with testing ZNS software > with QEMU. That is the most important goal here. 5.9 is around the corner, we > need something for people to get started with ZNS quickly. > Wait. What. No. Stop! It is unmistakably clear that you are invalidating my arguments about portability and endianness issues by suggesting that we just remove persistent state and deal with it later, but persistence is the killer feature that sets the QEMU emulated device apart from other emulation options. It is not about using emulation in production (because yeah, why would you?), but persistence is what makes it possible to develop and test "zoned FTLs" or something that requires recovery at power up. This is what allows testing of how your host software deals with opened zones being transitioned to FULL on power up and the persistent tracking of LBA allocation (in my series) can be used to properly test error recovery if you lost state in the app. Please, work with me on this instead of just removing such an essential feature. Since persistence seems to be the only thing we are really discussing, we should have plenty of time until the soft-freeze to come up with a proper solution on that. I agree that my version had a format that was pretty ad-hoc and that won't fly - it needs magic and version capabilities like in Dmitry's series, which incidentially looks a lot like what we did in the OpenChannel implementation, so I agree with the strategy. ZNS-wise, the only thing my implementation stores is the zone descriptors (in spec-native little-endian format) and the zone descriptor extensions. So there are no endian issues with those. The allocation tracking bitmap is always stored in little endian, but converted to big-endian if running on a big-endian host. Let me just conjure something up. #define NVME_PSTATE_MAGIC ... #define NVME_PSTATE_V1 1 typedef struct NvmePstateHeader { uint32_t magic; uint32_t version; uint64_t blk_len; uint8_t lbads; uint8_t iocs; uint8_t rsvd18[3054]; struct { uint64_t zsze; uint8_t zdes; } QEMU_PACKED zns; uint8_t rsvd3089[1007]; } QEMU_PACKED NvmePstateHeader; With such a header we have all we need. We can bail out if any parameters do not match and similar to nvme data structures it contains reserved areas for future use. I'll be posting a v2 with this. If this still feels too ad-hoc, we can be inspired by QCOW2 and the "extension" feature. I can agree that we drop other optional features like zone excursions and the reset/finish recommended limit simulation, but PLEASE DO NOT remove persistence and upstream a half-baked version when we are so close and have time to get it right.
On 2020/09/29 19:46, Klaus Jensen wrote: > On Sep 28 22:54, Damien Le Moal wrote: >> On 2020/09/29 6:25, Keith Busch wrote: >>> On Mon, Sep 28, 2020 at 08:36:48AM +0200, Klaus Jensen wrote: >>>> On Sep 28 02:33, Dmitry Fomichev wrote: >>>>> You are making it sound like the entire WDC series relies on this approach. >>>>> Actually, the persistency is introduced in the second to last patch in the >>>>> series and it only adds a couple of lines of code in the i/o path to mark >>>>> zones dirty. This is possible because of using mmap() and I find the way >>>>> it is done to be quite elegant, not ugly :) >>>>> >>>> >>>> No, I understand that your implementation works fine without >>>> persistance, but persistance is key. That is why my series adds it in >>>> the first patch. Without persistence it is just a toy. And the QEMU >>>> device is not just an "NVMe-version" of null_blk. >>> >>> I really think we should be a bit more cautious of commiting to an >>> on-disk format for the persistent state. Both this and Klaus' persistent >>> state feels a bit ad-hoc, and with all the other knobs provided, it >>> looks too easy to have out-of-sync states, or just not being able to >>> boot at all if a qemu versions have different on-disk formats. >>> >>> Is anyone really considering zone emulation for production level stuff >>> anyway? I can't imagine a real scenario where you'd want put yourself >>> through that: you are just giving yourself all the downsides of a zoned >>> block device and none of the benefits. AFAIK, this is provided as a >>> development vehicle, closer to a "toy". >>> >>> I think we should consider trimming this down to a more minimal set that >>> we *do* agree on and commit for inclusion ASAP. We can iterate all the >>> bells & whistles and flush out the meta data's data marshalling scheme >>> for persistence later. >> >> +1 on this. Removing the persistence also removes the debate on endianess. With >> that out of the way, it should be straightforward to get agreement on a series >> that can be merged quickly to get developers started with testing ZNS software >> with QEMU. That is the most important goal here. 5.9 is around the corner, we >> need something for people to get started with ZNS quickly. >> > > Wait. What. No. Stop! > > It is unmistakably clear that you are invalidating my arguments about > portability and endianness issues by suggesting that we just remove > persistent state and deal with it later, but persistence is the killer > feature that sets the QEMU emulated device apart from other emulation > options. It is not about using emulation in production (because yeah, > why would you?), but persistence is what makes it possible to develop > and test "zoned FTLs" or something that requires recovery at power up. > This is what allows testing of how your host software deals with opened > zones being transitioned to FULL on power up and the persistent tracking > of LBA allocation (in my series) can be used to properly test error > recovery if you lost state in the app. I am not invalidating anything. I am in violent agreement with you about the usefulness of persistence. My point was that I agree with Keith: let's first get the base emulation in and improve on top of it. And the base emulation does not need to include persistence and endianess of the saved zone meta for now. The result of this would still be super useful to have in stable. Then let's add persistence and others bells and whistles on top (see below). > Please, work with me on this instead of just removing such an essential > feature. Since persistence seems to be the only thing we are really > discussing, we should have plenty of time until the soft-freeze to come > up with a proper solution on that. > > I agree that my version had a format that was pretty ad-hoc and that > won't fly - it needs magic and version capabilities like in Dmitry's > series, which incidentially looks a lot like what we did in the > OpenChannel implementation, so I agree with the strategy. > > ZNS-wise, the only thing my implementation stores is the zone > descriptors (in spec-native little-endian format) and the zone > descriptor extensions. So there are no endian issues with those. The > allocation tracking bitmap is always stored in little endian, but > converted to big-endian if running on a big-endian host. > > Let me just conjure something up. > > #define NVME_PSTATE_MAGIC ... > #define NVME_PSTATE_V1 1 > > typedef struct NvmePstateHeader { > uint32_t magic; > uint32_t version; > > uint64_t blk_len; > > uint8_t lbads; > uint8_t iocs; > > uint8_t rsvd18[3054]; > > struct { > uint64_t zsze; > uint8_t zdes; > } QEMU_PACKED zns; > > uint8_t rsvd3089[1007]; > } QEMU_PACKED NvmePstateHeader; > > With such a header we have all we need. We can bail out if any > parameters do not match and similar to nvme data structures it contains > reserved areas for future use. I'll be posting a v2 with this. If this > still feels too ad-hoc, we can be inspired by QCOW2 and the "extension" > feature. > > I can agree that we drop other optional features like zone excursions > and the reset/finish recommended limit simulation, but PLEASE DO NOT > remove persistence and upstream a half-baked version when we are so > close and have time to get it right. OK. Then let's move the persistence implementation as the last patch in the series. This way, if it is still controversial, it will not block the rest. Here is what I propose: Dmitry: remove persistence stuff from your patches, address comments and resend. Klaus: Rebase your persistence patch(es) with reworked format on top of Dmitry series and send. That creates a pipeline for reviews and persistence is not a blocker. And I agree that other ZNS feature can come after we get all of that done first. Thoughts ? Keith ? Would that work for you ?
> -----Original Message----- > From: Klaus Jensen <its@irrelevant.dk> > Sent: Monday, September 28, 2020 2:37 AM > To: Dmitry Fomichev <Dmitry.Fomichev@wdc.com> > Cc: Keith Busch <kbusch@kernel.org>; Damien Le Moal > <Damien.LeMoal@wdc.com>; Klaus Jensen <k.jensen@samsung.com>; Kevin > Wolf <kwolf@redhat.com>; Philippe Mathieu-Daudé <philmd@redhat.com>; > Maxim Levitsky <mlevitsk@redhat.com>; Fam Zheng <fam@euphon.net>; > Niklas Cassel <Niklas.Cassel@wdc.com>; qemu-block@nongnu.org; qemu- > devel@nongnu.org; Alistair Francis <Alistair.Francis@wdc.com>; Matias > Bjorling <Matias.Bjorling@wdc.com> > Subject: Re: [PATCH v4 00/14] hw/block/nvme: Support Namespace Types > and Zoned Namespace Command Set > > On Sep 28 02:33, Dmitry Fomichev wrote: > > > -----Original Message----- > > > From: Klaus Jensen <its@irrelevant.dk> > > > > > > If it really needs to be memory mapped, then I think a hostmem-based > > > approach similar to what Andrzej did for PMR is needed (I think that > > > will get rid of the CONFIG_POSIX ifdef at least, but still leave it > > > slightly tricky to get it to work on all platforms AFAIK). > > > > Ok, it looks that using the HostMemoryBackendFile backend will be > > more appropriate. This will remove the need for conditional compile. > > > > The mmap() portability is pretty decent across software platforms. > > Any poor Windows user who is forced to emulate ZNS on mingw will be > > able to do so, just without having zone state persistency. Considering > > how specialized this stuff is in first place, I estimate the number of users > > affected by this "limitation" to be exactly zero. > > > > QEMU is a cross platform project - we should strive for portability. > > Alienating developers that use a Windows platform and calling them out > as "poor" is not exactly good for the zoned ecosystem. > Wow. By bringing up political correctness here you are basically admitting the fact that you have no real technical argument here. The whole Windows issue is red herring that you are using to attack the code that is absolutely legit, but comes from a competitor. Your initial complaint was that it doesn't compile in mingw and that it uses "wrong" API. You have even suggested the API to use. Now, the code uses that API and builds fine, but now it's still not good simply because you "do not like it". It's a disgrace. > > > But really, > > > since we do not require memory semantics for this, then I think the > > > abstraction is fundamentally wrong. > > > > > > > Seriously, what is wrong with using mmap :) ? It is used successfully for > > similar applications, for example - > > https://github.com/open-iscsi/tcmu-runner/blob/master/file_zbc.c > > > > There is nothing fundamentally wrong with mmap. I just think it is the > wrong abstraction here (and it limits portability for no good reason). > For PMR there is a good reason - it requires memory semantics. > We are trying to emulate NVMEe controller NVRAM. The best abstraction for emulating NVRAM would be... NVRAM! > > > I am, of course, blowing my own horn, since my implementation uses a > > > portable blockdev for this. > > > > > > > You are making it sound like the entire WDC series relies on this approach. > > Actually, the persistency is introduced in the second to last patch in the > > series and it only adds a couple of lines of code in the i/o path to mark > > zones dirty. This is possible because of using mmap() and I find the way > > it is done to be quite elegant, not ugly :) > > > > No, I understand that your implementation works fine without > persistance, but persistance is key. That is why my series adds it in > the first patch. Without persistence it is just a toy. And the QEMU > device is not just an "NVMe-version" of null_blk. > > And I don't think I ever called the use of mmap ugly. I called out the > physical memory API shenanigans as a hack. > > > > Another issue is the complete lack of endian conversions. Does it > > > matter? It depends. Will anyone ever use this on a big endian host and > > > move the meta data backing file to a little endian host? Probably not. > > > So does it really matter? Probably not, but it is cutting corners. > > > > > After I had replied this, I considered a follow-up, because there are > probably QEMU developers that would call me out on this. > > This definitely DOES matter to QEMU. > > > > > Great point on endianness! Naturally, all file backed values are stored in > > their native endianness. This way, there is no extra overhead on big endian > > hardware architectures. Portability concerns can be easily addressed by > > storing metadata endianness as a byte flag in its header. Then, during > > initialization, the metadata validation code can detect the possible > > discrepancy in endianness and automatically convert the metadata to the > > endianness of the host. This part is out of scope of this series, but I would > > be able to contribute such a solution as an enhancement in the future. > > > > It is not out of scope. I don't see why we should merge something that > is arguably buggy. Again, wow! Now you turned around and arbitrarily elevated this issue from moderate ("Does it matter?, cutting corners") to severe ("buggy"). Likely because v5 of WDC patchset has been posted. This, again, just shows your lack of integrity as a maintainer. This "issue" is a real trivial one to fix as I described above and you are blowing it up way out of proportion, making it look like it is a fundamental problem that can not be resolved. It's not. > > Bottomline is that I just don't see why we should accept an > implementation that > > a) excludes some platforms (Windows) from using persistence; and > b) contains endianness conversion issues > > when there is a portable implementation posted that at least tries to > convert endianness as needed. Doesn't that implementation discriminate against big endian architectures? :) Ok, it is a joke - with some folks I need to clarify this.
> -----Original Message----- > From: Qemu-block <qemu-block- > bounces+dmitry.fomichev=wdc.com@nongnu.org> On Behalf Of Klaus > Jensen > Sent: Tuesday, September 29, 2020 6:47 AM > To: Damien Le Moal <Damien.LeMoal@wdc.com> > Cc: Fam Zheng <fam@euphon.net>; Kevin Wolf <kwolf@redhat.com>; qemu- > block@nongnu.org; Niklas Cassel <Niklas.Cassel@wdc.com>; Klaus Jensen > <k.jensen@samsung.com>; qemu-devel@nongnu.org; Alistair Francis > <Alistair.Francis@wdc.com>; Keith Busch <kbusch@kernel.org>; Philippe > Mathieu-Daudé <philmd@redhat.com>; Matias Bjorling > <Matias.Bjorling@wdc.com> > Subject: Re: [PATCH v4 00/14] hw/block/nvme: Support Namespace Types > and Zoned Namespace Command Set > > On Sep 28 22:54, Damien Le Moal wrote: > > On 2020/09/29 6:25, Keith Busch wrote: > > > On Mon, Sep 28, 2020 at 08:36:48AM +0200, Klaus Jensen wrote: > > >> On Sep 28 02:33, Dmitry Fomichev wrote: > > >>> You are making it sound like the entire WDC series relies on this > approach. > > >>> Actually, the persistency is introduced in the second to last patch in the > > >>> series and it only adds a couple of lines of code in the i/o path to mark > > >>> zones dirty. This is possible because of using mmap() and I find the way > > >>> it is done to be quite elegant, not ugly :) > > >>> > > >> > > >> No, I understand that your implementation works fine without > > >> persistance, but persistance is key. That is why my series adds it in > > >> the first patch. Without persistence it is just a toy. And the QEMU > > >> device is not just an "NVMe-version" of null_blk. > > > > > > I really think we should be a bit more cautious of commiting to an > > > on-disk format for the persistent state. Both this and Klaus' persistent > > > state feels a bit ad-hoc, and with all the other knobs provided, it > > > looks too easy to have out-of-sync states, or just not being able to > > > boot at all if a qemu versions have different on-disk formats. > > > > > > Is anyone really considering zone emulation for production level stuff > > > anyway? I can't imagine a real scenario where you'd want put yourself > > > through that: you are just giving yourself all the downsides of a zoned > > > block device and none of the benefits. AFAIK, this is provided as a > > > development vehicle, closer to a "toy". > > > > > > I think we should consider trimming this down to a more minimal set that > > > we *do* agree on and commit for inclusion ASAP. We can iterate all the > > > bells & whistles and flush out the meta data's data marshalling scheme > > > for persistence later. > > > > +1 on this. Removing the persistence also removes the debate on > endianess. With > > that out of the way, it should be straightforward to get agreement on a > series > > that can be merged quickly to get developers started with testing ZNS > software > > with QEMU. That is the most important goal here. 5.9 is around the corner, > we > > need something for people to get started with ZNS quickly. > > > > Wait. What. No. Stop! > > It is unmistakably clear that you are invalidating my arguments about > portability and endianness issues by suggesting that we just remove > persistent state and deal with it later, but persistence is the killer > feature that sets the QEMU emulated device apart from other emulation > options. It is not about using emulation in production (because yeah, > why would you?), but persistence is what makes it possible to develop > and test "zoned FTLs" or something that requires recovery at power up. > This is what allows testing of how your host software deals with opened > zones being transitioned to FULL on power up and the persistent tracking > of LBA allocation (in my series) can be used to properly test error > recovery if you lost state in the app. > > Please, work with me on this instead of just removing such an essential > feature. Since persistence seems to be the only thing we are really > discussing, we should have plenty of time until the soft-freeze to come > up with a proper solution on that. > > I agree that my version had a format that was pretty ad-hoc and that > won't fly - it needs magic and version capabilities like in Dmitry's > series, which incidentially looks a lot like what we did in the > OpenChannel implementation, so I agree with the strategy. Are you insinuating that I somehow took stuff from OCSSD code and try to claim priority this way? I am not at all that familiar with that code. And I've already sent you the link to tcmu-runner code that served me as an inspiration for implementing persistence in WDC patchset. That code has been around for years, uses mmap, works great and has nothing to do with you. > > ZNS-wise, the only thing my implementation stores is the zone > descriptors (in spec-native little-endian format) and the zone > descriptor extensions. So there are no endian issues with those. The > allocation tracking bitmap is always stored in little endian, but > converted to big-endian if running on a big-endian host. > > Let me just conjure something up. > > #define NVME_PSTATE_MAGIC ... > #define NVME_PSTATE_V1 1 > > typedef struct NvmePstateHeader { > uint32_t magic; > uint32_t version; > > uint64_t blk_len; > > uint8_t lbads; > uint8_t iocs; > > uint8_t rsvd18[3054]; > > struct { > uint64_t zsze; > uint8_t zdes; > } QEMU_PACKED zns; > > uint8_t rsvd3089[1007]; > } QEMU_PACKED NvmePstateHeader; > Why conjure something that already exists in WDC patchset? And that part has been published in the very first version of our patches, weeks before your entire ZNS series was posted. Add an rsvd[] here and there and that code can be as suitable to achieve the stated goals as what you have above. > With such a header we have all we need. We can bail out if any > parameters do not match and similar to nvme data structures it contains > reserved areas for future use. I'll be posting a v2 with this. If this > still feels too ad-hoc, we can be inspired by QCOW2 and the "extension" > feature. > > I can agree that we drop other optional features like zone excursions > and the reset/finish recommended limit simulation, but PLEASE DO NOT > remove persistence and upstream a half-baked version when we are so > close and have time to get it right. One important thing IMO is to reduce future need for metadata versioning. This requires a really good effort to design and review the proper metadata format that would become stable for some time. Think about portability. To get out something "conjured up" now and then have to move to V2 metadata in the next release is even worse than no persistence at all. So maybe is makes sense to go with Keith's suggestion.
On Sep 29 15:43, Dmitry Fomichev wrote: > > -----Original Message----- > > From: Qemu-block <qemu-block- > > bounces+dmitry.fomichev=wdc.com@nongnu.org> On Behalf Of Klaus > > Jensen > > Sent: Tuesday, September 29, 2020 6:47 AM > > To: Damien Le Moal <Damien.LeMoal@wdc.com> > > Cc: Fam Zheng <fam@euphon.net>; Kevin Wolf <kwolf@redhat.com>; qemu- > > block@nongnu.org; Niklas Cassel <Niklas.Cassel@wdc.com>; Klaus Jensen > > <k.jensen@samsung.com>; qemu-devel@nongnu.org; Alistair Francis > > <Alistair.Francis@wdc.com>; Keith Busch <kbusch@kernel.org>; Philippe > > Mathieu-Daudé <philmd@redhat.com>; Matias Bjorling > > <Matias.Bjorling@wdc.com> > > Subject: Re: [PATCH v4 00/14] hw/block/nvme: Support Namespace Types > > and Zoned Namespace Command Set > > > > On Sep 28 22:54, Damien Le Moal wrote: > > > On 2020/09/29 6:25, Keith Busch wrote: > > > > On Mon, Sep 28, 2020 at 08:36:48AM +0200, Klaus Jensen wrote: > > > >> On Sep 28 02:33, Dmitry Fomichev wrote: > > > >>> You are making it sound like the entire WDC series relies on this > > approach. > > > >>> Actually, the persistency is introduced in the second to last patch in the > > > >>> series and it only adds a couple of lines of code in the i/o path to mark > > > >>> zones dirty. This is possible because of using mmap() and I find the way > > > >>> it is done to be quite elegant, not ugly :) > > > >>> > > > >> > > > >> No, I understand that your implementation works fine without > > > >> persistance, but persistance is key. That is why my series adds it in > > > >> the first patch. Without persistence it is just a toy. And the QEMU > > > >> device is not just an "NVMe-version" of null_blk. > > > > > > > > I really think we should be a bit more cautious of commiting to an > > > > on-disk format for the persistent state. Both this and Klaus' persistent > > > > state feels a bit ad-hoc, and with all the other knobs provided, it > > > > looks too easy to have out-of-sync states, or just not being able to > > > > boot at all if a qemu versions have different on-disk formats. > > > > > > > > Is anyone really considering zone emulation for production level stuff > > > > anyway? I can't imagine a real scenario where you'd want put yourself > > > > through that: you are just giving yourself all the downsides of a zoned > > > > block device and none of the benefits. AFAIK, this is provided as a > > > > development vehicle, closer to a "toy". > > > > > > > > I think we should consider trimming this down to a more minimal set that > > > > we *do* agree on and commit for inclusion ASAP. We can iterate all the > > > > bells & whistles and flush out the meta data's data marshalling scheme > > > > for persistence later. > > > > > > +1 on this. Removing the persistence also removes the debate on > > endianess. With > > > that out of the way, it should be straightforward to get agreement on a > > series > > > that can be merged quickly to get developers started with testing ZNS > > software > > > with QEMU. That is the most important goal here. 5.9 is around the corner, > > we > > > need something for people to get started with ZNS quickly. > > > > > > > Wait. What. No. Stop! > > > > It is unmistakably clear that you are invalidating my arguments about > > portability and endianness issues by suggesting that we just remove > > persistent state and deal with it later, but persistence is the killer > > feature that sets the QEMU emulated device apart from other emulation > > options. It is not about using emulation in production (because yeah, > > why would you?), but persistence is what makes it possible to develop > > and test "zoned FTLs" or something that requires recovery at power up. > > This is what allows testing of how your host software deals with opened > > zones being transitioned to FULL on power up and the persistent tracking > > of LBA allocation (in my series) can be used to properly test error > > recovery if you lost state in the app. > > > > Please, work with me on this instead of just removing such an essential > > feature. Since persistence seems to be the only thing we are really > > discussing, we should have plenty of time until the soft-freeze to come > > up with a proper solution on that. > > > > I agree that my version had a format that was pretty ad-hoc and that > > won't fly - it needs magic and version capabilities like in Dmitry's > > series, which incidentially looks a lot like what we did in the > > OpenChannel implementation, so I agree with the strategy. > > Are you insinuating that I somehow took stuff from OCSSD code and try > to claim priority this way? I am not at all that familiar with that code. > And I've already sent you the link to tcmu-runner code that served me > as an inspiration for implementing persistence in WDC patchset. > That code has been around for years, uses mmap, works great and has > nothing to do with you. > No. I am not insinuating anything. The OpenChannel device also used a blockdev, but, yes, incidentially (and sorry, I should not have used that word), it looked like how we did it there and I noted that I agreed with the strategy. > > > > ZNS-wise, the only thing my implementation stores is the zone > > descriptors (in spec-native little-endian format) and the zone > > descriptor extensions. So there are no endian issues with those. The > > allocation tracking bitmap is always stored in little endian, but > > converted to big-endian if running on a big-endian host. > > > > Let me just conjure something up. > > > > #define NVME_PSTATE_MAGIC ... > > #define NVME_PSTATE_V1 1 > > > > typedef struct NvmePstateHeader { > > uint32_t magic; > > uint32_t version; > > > > uint64_t blk_len; > > > > uint8_t lbads; > > uint8_t iocs; > > > > uint8_t rsvd18[3054]; > > > > struct { > > uint64_t zsze; > > uint8_t zdes; > > } QEMU_PACKED zns; > > > > uint8_t rsvd3089[1007]; > > } QEMU_PACKED NvmePstateHeader; > > > > Why conjure something that already exists in WDC patchset? And that part > has been published in the very first version of our patches, weeks before > your entire ZNS series was posted. Add an rsvd[] here and there and that code > can be as suitable to achieve the stated goals as what you have above. > Yes, I read your code. I know you have a header and I also noted above that "it needs magic and version capabilities like in Dmitry's series". > > series, > > With such a header we have all we need. We can bail out if any > > parameters do not match and similar to nvme data structures it contains > > reserved areas for future use. I'll be posting a v2 with this. If this > > still feels too ad-hoc, we can be inspired by QCOW2 and the "extension" > > feature. > > > > I can agree that we drop other optional features like zone excursions > > and the reset/finish recommended limit simulation, but PLEASE DO NOT > > remove persistence and upstream a half-baked version when we are so > > close and have time to get it right. > > One important thing IMO is to reduce future need for metadata versioning. > This requires a really good effort to design and review the proper metadata > format that would become stable for some time. Think about portability. > To get out something "conjured up" now and then have to move to V2 > metadata in the next release is even worse than no persistence at all. > So maybe is makes sense to go with Keith's suggestion. As I've said, we have time until the soft-freeze to get this right. I "conjured" something up to show a point. The reason we review and iterate is to NOT upstream something thats is conjured up. But we gotta start somewhere, no? So what is so bad about me posting a v2?
On Tue, Sep 29, 2020 at 12:46:33PM +0200, Klaus Jensen wrote: > It is unmistakably clear that you are invalidating my arguments about > portability and endianness issues by suggesting that we just remove > persistent state and deal with it later, but persistence is the killer > feature that sets the QEMU emulated device apart from other emulation > options. It is not about using emulation in production (because yeah, > why would you?), but persistence is what makes it possible to develop > and test "zoned FTLs" or something that requires recovery at power up. > This is what allows testing of how your host software deals with opened > zones being transitioned to FULL on power up and the persistent tracking > of LBA allocation (in my series) can be used to properly test error > recovery if you lost state in the app. Hold up -- why does an OPEN zone transition to FULL on power up? The spec suggests it should be CLOSED. The spec does appear to support going to FULL on a NVM Subsystem Reset, though. Actually, now that I'm looking at this part of the spec, these implicit transitions seem a bit less clear than I expected. I'm not sure it's clear enough to evaluate qemu's compliance right now. But I don't see what testing these transitions has to do with having a persistent state. You can reboot your VM without tearing down the running QEMU instance. You can also unbind the driver or shutdown the controller within the running operating system. That should make those implicit state transitions reachable in order to exercise your FTL's recovery. I agree the persistent state provides conveniences for developers. I just don't want to gate ZNS enabling on it either since the core design doesn't depend on it.
On Tue, Sep 29, 2020 at 11:13:51AM +0000, Damien Le Moal wrote: > OK. Then let's move the persistence implementation as the last patch in the > series. This way, if it is still controversial, it will not block the rest. > > Here is what I propose: > Dmitry: remove persistence stuff from your patches, address comments and resend. > Klaus: Rebase your persistence patch(es) with reworked format on top of Dmitry > series and send. > > That creates a pipeline for reviews and persistence is not a blocker. And I > agree that other ZNS feature can come after we get all of that done first. > > Thoughts ? Keith ? Would that work for you ? That works for me. I will have comments for Dmitry's v5, though, so please wait one more day before considering a respin.
On Sep 29 10:29, Keith Busch wrote: > On Tue, Sep 29, 2020 at 12:46:33PM +0200, Klaus Jensen wrote: > > It is unmistakably clear that you are invalidating my arguments about > > portability and endianness issues by suggesting that we just remove > > persistent state and deal with it later, but persistence is the killer > > feature that sets the QEMU emulated device apart from other emulation > > options. It is not about using emulation in production (because yeah, > > why would you?), but persistence is what makes it possible to develop > > and test "zoned FTLs" or something that requires recovery at power up. > > This is what allows testing of how your host software deals with opened > > zones being transitioned to FULL on power up and the persistent tracking > > of LBA allocation (in my series) can be used to properly test error > > recovery if you lost state in the app. > > Hold up -- why does an OPEN zone transition to FULL on power up? The > spec suggests it should be CLOSED. The spec does appear to support going > to FULL on a NVM Subsystem Reset, though. Actually, now that I'm looking > at this part of the spec, these implicit transitions seem a bit less > clear than I expected. I'm not sure it's clear enough to evaluate qemu's > compliance right now. > > But I don't see what testing these transitions has to do with having a > persistent state. You can reboot your VM without tearing down the > running QEMU instance. You can also unbind the driver or shutdown the > controller within the running operating system. That should make those > implicit state transitions reachable in order to exercise your FTL's > recovery. > Oh dear - don't "spec" with me ;) NVMe v1.4 Section 7.3.1: An NVM Subsystem Reset is initiated when: * Main power is applied to the NVM subsystem; * A value of 4E564D64h ("NVMe") is written to the NSSR.NSSRC field; * Requested using a method defined in the NVMe Management Interface specification; or * A vendor specific event occurs. In the context of QEMU, "Main power" is tearing down QEMU and starting it from scratch. Just like on a "real" host, unbinding the driver, rebooting or shutting down the controller does not cause a subsystem reset (and does not cause the zones to change state). And since the device does not indicate support for the optional NSSR.NSSRC register, that way to initiate a subsystem cannot be used. The reason for moving to FULL is that write pointer updates are not persisted on each advancement, only when the zone state changes. So zones that were opened might have valid data, but invalid write pointer. So the device transitions them to FULL as it is allowed to. QED. > I agree the persistent state provides conveniences for developers. I > just don't want to gate ZNS enabling on it either since the core design > doesn't depend on it. I just don't see why we cant have the icing on the cake when it is already there :)
On Tue, Sep 29, 2020 at 08:00:04PM +0200, Klaus Jensen wrote: > On Sep 29 10:29, Keith Busch wrote: > > On Tue, Sep 29, 2020 at 12:46:33PM +0200, Klaus Jensen wrote: > > > It is unmistakably clear that you are invalidating my arguments about > > > portability and endianness issues by suggesting that we just remove > > > persistent state and deal with it later, but persistence is the killer > > > feature that sets the QEMU emulated device apart from other emulation > > > options. It is not about using emulation in production (because yeah, > > > why would you?), but persistence is what makes it possible to develop > > > and test "zoned FTLs" or something that requires recovery at power up. > > > This is what allows testing of how your host software deals with opened > > > zones being transitioned to FULL on power up and the persistent tracking > > > of LBA allocation (in my series) can be used to properly test error > > > recovery if you lost state in the app. > > > > Hold up -- why does an OPEN zone transition to FULL on power up? The > > spec suggests it should be CLOSED. The spec does appear to support going > > to FULL on a NVM Subsystem Reset, though. Actually, now that I'm looking > > at this part of the spec, these implicit transitions seem a bit less > > clear than I expected. I'm not sure it's clear enough to evaluate qemu's > > compliance right now. > > > > But I don't see what testing these transitions has to do with having a > > persistent state. You can reboot your VM without tearing down the > > running QEMU instance. You can also unbind the driver or shutdown the > > controller within the running operating system. That should make those > > implicit state transitions reachable in order to exercise your FTL's > > recovery. > > > > Oh dear - don't "spec" with me ;) > > NVMe v1.4 Section 7.3.1: > > An NVM Subsystem Reset is initiated when: > * Main power is applied to the NVM subsystem; > * A value of 4E564D64h ("NVMe") is written to the NSSR.NSSRC > field; > * Requested using a method defined in the NVMe Management > Interface specification; or > * A vendor specific event occurs. Okay. I wish the nvme twg would strip the changelog from the published TPs. We have unhelpful statements like this in the ZNS spec: "Default active zones to transition to Closed state on power/controller reset." > In the context of QEMU, "Main power" is tearing down QEMU and starting > it from scratch. Just like on a "real" host, unbinding the driver, > rebooting or shutting down the controller does not cause a subsystem > reset (and does not cause the zones to change state). That can't be right. The ZNS spec says: The initial state of a zone state machine is set as a result of: a) an NVM Subsystem Reset; or b) all controllers in the NVM subsystem reporting Shutdown processing complete ((i.e., 10b in the Shutdown Status (SHST) register) So a CC.SHN had better cause an implicit transition of open zones to their "initial" state since 'open' is not a valid initial state.
> -----Original Message----- > From: Klaus Jensen <its@irrelevant.dk> > Sent: Tuesday, 29 September 2020 20.00 > To: Keith Busch <kbusch@kernel.org> > Cc: Damien Le Moal <Damien.LeMoal@wdc.com>; Fam Zheng > <fam@euphon.net>; Kevin Wolf <kwolf@redhat.com>; qemu- > block@nongnu.org; Niklas Cassel <Niklas.Cassel@wdc.com>; Klaus Jensen > <k.jensen@samsung.com>; qemu-devel@nongnu.org; Alistair Francis > <Alistair.Francis@wdc.com>; Philippe Mathieu-Daudé <philmd@redhat.com>; > Matias Bjorling <Matias.Bjorling@wdc.com> > Subject: Re: [PATCH v4 00/14] hw/block/nvme: Support Namespace Types and > Zoned Namespace Command Set > > On Sep 29 10:29, Keith Busch wrote: > > On Tue, Sep 29, 2020 at 12:46:33PM +0200, Klaus Jensen wrote: > > > It is unmistakably clear that you are invalidating my arguments > > > about portability and endianness issues by suggesting that we just > > > remove persistent state and deal with it later, but persistence is > > > the killer feature that sets the QEMU emulated device apart from > > > other emulation options. It is not about using emulation in > > > production (because yeah, why would you?), but persistence is what > > > makes it possible to develop and test "zoned FTLs" or something that > requires recovery at power up. > > > This is what allows testing of how your host software deals with > > > opened zones being transitioned to FULL on power up and the > > > persistent tracking of LBA allocation (in my series) can be used to > > > properly test error recovery if you lost state in the app. > > > > Hold up -- why does an OPEN zone transition to FULL on power up? The > > spec suggests it should be CLOSED. The spec does appear to support > > going to FULL on a NVM Subsystem Reset, though. Actually, now that I'm > > looking at this part of the spec, these implicit transitions seem a > > bit less clear than I expected. I'm not sure it's clear enough to > > evaluate qemu's compliance right now. > > > > But I don't see what testing these transitions has to do with having a > > persistent state. You can reboot your VM without tearing down the > > running QEMU instance. You can also unbind the driver or shutdown the > > controller within the running operating system. That should make those > > implicit state transitions reachable in order to exercise your FTL's > > recovery. > > > > Oh dear - don't "spec" with me ;) > > NVMe v1.4 Section 7.3.1: > > An NVM Subsystem Reset is initiated when: > * Main power is applied to the NVM subsystem; > * A value of 4E564D64h ("NVMe") is written to the NSSR.NSSRC > field; > * Requested using a method defined in the NVMe Management > Interface specification; or > * A vendor specific event occurs. > > In the context of QEMU, "Main power" is tearing down QEMU and starting it > from scratch. Just like on a "real" host, unbinding the driver, rebooting or > shutting down the controller does not cause a subsystem reset (and does not > cause the zones to change state). And since the device does not indicate > support for the optional NSSR.NSSRC register, that way to initiate a subsystem > cannot be used. > > The reason for moving to FULL is that write pointer updates are not persisted > on each advancement, only when the zone state changes. So zones that were > opened might have valid data, but invalid write pointer. > So the device transitions them to FULL as it is allowed to. > How about when one must also recover from intermediate states (i.e., open/closed upon power loss). For example, I don't hope a real SSD implementation transition zones to full when it has thousands of open simultaneously. That could be a disaster for the PE cycles, and a lot of media going to waste. One would want applications to support that kind of failure mode as well.
On Sep 29 11:15, Keith Busch wrote: > On Tue, Sep 29, 2020 at 08:00:04PM +0200, Klaus Jensen wrote: > > On Sep 29 10:29, Keith Busch wrote: > > > On Tue, Sep 29, 2020 at 12:46:33PM +0200, Klaus Jensen wrote: > > > > It is unmistakably clear that you are invalidating my arguments about > > > > portability and endianness issues by suggesting that we just remove > > > > persistent state and deal with it later, but persistence is the killer > > > > feature that sets the QEMU emulated device apart from other emulation > > > > options. It is not about using emulation in production (because yeah, > > > > why would you?), but persistence is what makes it possible to develop > > > > and test "zoned FTLs" or something that requires recovery at power up. > > > > This is what allows testing of how your host software deals with opened > > > > zones being transitioned to FULL on power up and the persistent tracking > > > > of LBA allocation (in my series) can be used to properly test error > > > > recovery if you lost state in the app. > > > > > > Hold up -- why does an OPEN zone transition to FULL on power up? The > > > spec suggests it should be CLOSED. The spec does appear to support going > > > to FULL on a NVM Subsystem Reset, though. Actually, now that I'm looking > > > at this part of the spec, these implicit transitions seem a bit less > > > clear than I expected. I'm not sure it's clear enough to evaluate qemu's > > > compliance right now. > > > > > > But I don't see what testing these transitions has to do with having a > > > persistent state. You can reboot your VM without tearing down the > > > running QEMU instance. You can also unbind the driver or shutdown the > > > controller within the running operating system. That should make those > > > implicit state transitions reachable in order to exercise your FTL's > > > recovery. > > > > > > > Oh dear - don't "spec" with me ;) > > > > NVMe v1.4 Section 7.3.1: > > > > An NVM Subsystem Reset is initiated when: > > * Main power is applied to the NVM subsystem; > > * A value of 4E564D64h ("NVMe") is written to the NSSR.NSSRC > > field; > > * Requested using a method defined in the NVMe Management > > Interface specification; or > > * A vendor specific event occurs. > > Okay. I wish the nvme twg would strip the changelog from the published > TPs. We have unhelpful statements like this in the ZNS spec: > > "Default active zones to transition to Closed state on power/controller reset." > > > In the context of QEMU, "Main power" is tearing down QEMU and starting > > it from scratch. Just like on a "real" host, unbinding the driver, > > rebooting or shutting down the controller does not cause a subsystem > > reset (and does not cause the zones to change state). > > That can't be right. The ZNS spec says: > > The initial state of a zone state machine is set as a result of: > a) an NVM Subsystem Reset; or > b) all controllers in the NVM subsystem reporting Shutdown > processing complete ((i.e., 10b in the Shutdown Status (SHST) > register) > > So a CC.SHN had better cause an implicit transition of open zones to > their "initial" state since 'open' is not a valid initial state. Oh snap; true, you got me there.
On Sep 29 18:17, Matias Bjorling wrote: > > > > -----Original Message----- > > From: Klaus Jensen <its@irrelevant.dk> > > Sent: Tuesday, 29 September 2020 20.00 > > To: Keith Busch <kbusch@kernel.org> > > Cc: Damien Le Moal <Damien.LeMoal@wdc.com>; Fam Zheng > > <fam@euphon.net>; Kevin Wolf <kwolf@redhat.com>; qemu- > > block@nongnu.org; Niklas Cassel <Niklas.Cassel@wdc.com>; Klaus Jensen > > <k.jensen@samsung.com>; qemu-devel@nongnu.org; Alistair Francis > > <Alistair.Francis@wdc.com>; Philippe Mathieu-Daudé <philmd@redhat.com>; > > Matias Bjorling <Matias.Bjorling@wdc.com> > > Subject: Re: [PATCH v4 00/14] hw/block/nvme: Support Namespace Types and > > Zoned Namespace Command Set > > > > On Sep 29 10:29, Keith Busch wrote: > > > On Tue, Sep 29, 2020 at 12:46:33PM +0200, Klaus Jensen wrote: > > > > It is unmistakably clear that you are invalidating my arguments > > > > about portability and endianness issues by suggesting that we just > > > > remove persistent state and deal with it later, but persistence is > > > > the killer feature that sets the QEMU emulated device apart from > > > > other emulation options. It is not about using emulation in > > > > production (because yeah, why would you?), but persistence is what > > > > makes it possible to develop and test "zoned FTLs" or something that > > requires recovery at power up. > > > > This is what allows testing of how your host software deals with > > > > opened zones being transitioned to FULL on power up and the > > > > persistent tracking of LBA allocation (in my series) can be used to > > > > properly test error recovery if you lost state in the app. > > > > > > Hold up -- why does an OPEN zone transition to FULL on power up? The > > > spec suggests it should be CLOSED. The spec does appear to support > > > going to FULL on a NVM Subsystem Reset, though. Actually, now that I'm > > > looking at this part of the spec, these implicit transitions seem a > > > bit less clear than I expected. I'm not sure it's clear enough to > > > evaluate qemu's compliance right now. > > > > > > But I don't see what testing these transitions has to do with having a > > > persistent state. You can reboot your VM without tearing down the > > > running QEMU instance. You can also unbind the driver or shutdown the > > > controller within the running operating system. That should make those > > > implicit state transitions reachable in order to exercise your FTL's > > > recovery. > > > > > > > Oh dear - don't "spec" with me ;) > > > > NVMe v1.4 Section 7.3.1: > > > > An NVM Subsystem Reset is initiated when: > > * Main power is applied to the NVM subsystem; > > * A value of 4E564D64h ("NVMe") is written to the NSSR.NSSRC > > field; > > * Requested using a method defined in the NVMe Management > > Interface specification; or > > * A vendor specific event occurs. > > > > In the context of QEMU, "Main power" is tearing down QEMU and starting it > > from scratch. Just like on a "real" host, unbinding the driver, rebooting or > > shutting down the controller does not cause a subsystem reset (and does not > > cause the zones to change state). And since the device does not indicate > > support for the optional NSSR.NSSRC register, that way to initiate a subsystem > > cannot be used. > > > > The reason for moving to FULL is that write pointer updates are not persisted > > on each advancement, only when the zone state changes. So zones that were > > opened might have valid data, but invalid write pointer. > > So the device transitions them to FULL as it is allowed to. > > > > How about when one must also recover from intermediate states (i.e., > open/closed upon power loss). For example, I don't hope a real SSD > implementation transition zones to full when it has thousands of open > simultaneously. That could be a disaster for the PE cycles, and a lot > of media going to waste. One would want applications to support that > kind of failure mode as well. Christ. The WDC Strike Force is really jumping out of lightspeed here. I'm afraid I don't have an opposing force to engage with. So I'll be your only boxing bag for the evening. As Keith just said, "Opened" is not a valid intial state. Didn't you write the spec? ;) As for Closed, they will be brought up as is. With that in mind, I'm not sure what you specifically refer to? I'll gently remind you that the QEMU nvme device is not a real SSD and does not deal with NAND so it does not really do any "recovering" of intermediate states on power on if that is what you refer to?
On Sep 29 15:42, Dmitry Fomichev wrote: > > -----Original Message----- > > From: Klaus Jensen <its@irrelevant.dk> > > Sent: Monday, September 28, 2020 2:37 AM > > To: Dmitry Fomichev <Dmitry.Fomichev@wdc.com> > > Cc: Keith Busch <kbusch@kernel.org>; Damien Le Moal > > <Damien.LeMoal@wdc.com>; Klaus Jensen <k.jensen@samsung.com>; Kevin > > Wolf <kwolf@redhat.com>; Philippe Mathieu-Daudé <philmd@redhat.com>; > > Maxim Levitsky <mlevitsk@redhat.com>; Fam Zheng <fam@euphon.net>; > > Niklas Cassel <Niklas.Cassel@wdc.com>; qemu-block@nongnu.org; qemu- > > devel@nongnu.org; Alistair Francis <Alistair.Francis@wdc.com>; Matias > > Bjorling <Matias.Bjorling@wdc.com> > > Subject: Re: [PATCH v4 00/14] hw/block/nvme: Support Namespace Types > > and Zoned Namespace Command Set > > > > On Sep 28 02:33, Dmitry Fomichev wrote: > > > > -----Original Message----- > > > > From: Klaus Jensen <its@irrelevant.dk> > > > > > > > > If it really needs to be memory mapped, then I think a hostmem-based > > > > approach similar to what Andrzej did for PMR is needed (I think that > > > > will get rid of the CONFIG_POSIX ifdef at least, but still leave it > > > > slightly tricky to get it to work on all platforms AFAIK). > > > > > > Ok, it looks that using the HostMemoryBackendFile backend will be > > > more appropriate. This will remove the need for conditional compile. > > > > > > The mmap() portability is pretty decent across software platforms. > > > Any poor Windows user who is forced to emulate ZNS on mingw will be > > > able to do so, just without having zone state persistency. Considering > > > how specialized this stuff is in first place, I estimate the number of users > > > affected by this "limitation" to be exactly zero. > > > > > > > QEMU is a cross platform project - we should strive for portability. > > > > Alienating developers that use a Windows platform and calling them out > > as "poor" is not exactly good for the zoned ecosystem. > > > > Wow. By bringing up political correctness here you are basically admitting > the fact that you have no real technical argument here. I prefer that we support all platforms if and when we can. That's a technical argument, not a personal one like you those you start using now. > The whole Windows issue is red herring that you are using to attack > the code that is absolutely legit, but comes from a competitor. I can't even... > Your initial complaint was that it doesn't compile in mingw and that > it uses "wrong" API. You have even suggested the API to use. Now, the > code uses that API and builds fine, but now it's still not good simply > because you "do not like it". It's a disgrace. > I answered this in a previous reply. > > > > But really, > > > > since we do not require memory semantics for this, then I think the > > > > abstraction is fundamentally wrong. > > > > > > > > > > Seriously, what is wrong with using mmap :) ? It is used successfully for > > > similar applications, for example - > > > https://github.com/open-iscsi/tcmu-runner/blob/master/file_zbc.c > > > > > > > There is nothing fundamentally wrong with mmap. I just think it is the > > wrong abstraction here (and it limits portability for no good reason). > > For PMR there is a good reason - it requires memory semantics. > > > > We are trying to emulate NVMEe controller NVRAM. The best abstraction > for emulating NVRAM would be... NVRAM! > You never brought that up before and sure it could be a fair argument, except it is not true. PMR is emulating NVRAM (and requires memory semantics). Persistent state is not emulating anything. It is an implementation detail. > > > > I am, of course, blowing my own horn, since my implementation uses a > > > > portable blockdev for this. > > > > > > > > > > You are making it sound like the entire WDC series relies on this approach. > > > Actually, the persistency is introduced in the second to last patch in the > > > series and it only adds a couple of lines of code in the i/o path to mark > > > zones dirty. This is possible because of using mmap() and I find the way > > > it is done to be quite elegant, not ugly :) > > > > > > > No, I understand that your implementation works fine without > > persistance, but persistance is key. That is why my series adds it in > > the first patch. Without persistence it is just a toy. And the QEMU > > device is not just an "NVMe-version" of null_blk. > > > > And I don't think I ever called the use of mmap ugly. I called out the > > physical memory API shenanigans as a hack. > > > > > > Another issue is the complete lack of endian conversions. Does it > > > > matter? It depends. Will anyone ever use this on a big endian host and > > > > move the meta data backing file to a little endian host? Probably not. > > > > So does it really matter? Probably not, but it is cutting corners. > > > > > > > > After I had replied this, I considered a follow-up, because there are > > probably QEMU developers that would call me out on this. > > > > This definitely DOES matter to QEMU. > > > > > > > > Great point on endianness! Naturally, all file backed values are stored in > > > their native endianness. This way, there is no extra overhead on big endian > > > hardware architectures. Portability concerns can be easily addressed by > > > storing metadata endianness as a byte flag in its header. Then, during > > > initialization, the metadata validation code can detect the possible > > > discrepancy in endianness and automatically convert the metadata to the > > > endianness of the host. This part is out of scope of this series, but I would > > > be able to contribute such a solution as an enhancement in the future. > > > > > > > It is not out of scope. I don't see why we should merge something that > > is arguably buggy. > > Again, wow! Now you turned around and arbitrarily elevated this issue from > moderate ("Does it matter?, cutting corners") to severe ("buggy"). Likely > because v5 of WDC patchset has been posted. No, exactly as I wrote above, after I hit reply I considered a follow-up. I guess I should have. > This, again, just shows your lack of integrity as a maintainer. > I can't believe I just read that. I will not put up with this. It is completely non-called for. I stand up for my opinions and I will fight to make sure the best possible code goes upstream. Yes, I am paid by Samsung. But I can compartmentalize. I have been working on QEMU before Samsung and I know how to separate corporate interest and open source. I have a proven record on this list to show that. I really cannot believe that you brought it down to that level. I have been putting forth technical arguments throughout this entire review process and now you are getting personal. Not. Cool. Please keep things professional from now. > This "issue" is a real trivial one to fix as I described above and you are > blowing it up way out of proportion, making it look like it is a fundamental > problem that can not be resolved. It's not. > If it is so trival to fix, please fix it. I think I made it clear that I won't be happy until it is portable. And please note that I have *not* complained about other parts of your series. I have complained ALOT about the persistence implementation - and I continue to stand behind those complaints. I'm getting super tired of this one-sided process. I have continuously reviewed and commented your series, I have found multiple bugs, I have suggested improvements. Maybe if just one or two of those 9 people who signed off on your zoned implementation could look past their own nose and look at my series - you might just realize that its decent, portable and offers some niceties that yours do not have (at the cost of the same amount of code mind you).
All, Let's de-escalate this, please. There's no reason to doubt Klaus wants to see this to work well, just as everyone else does. We unfortunately have conflicting proposals posted, and everyone is passionate enough about their work, but please simmer down. As I mentioned earlier, I'd like to refocus on the basic implementation and save the persistent state discussion once the core is solid. After going through it all, I feel there's enough to discuss there to keep us busy for little while longer. Additional comments on the code will be coming from me later today.
> -----Original Message----- > From: Klaus Jensen <its@irrelevant.dk> > Sent: Tuesday, 29 September 2020 20.36 > To: Matias Bjorling <Matias.Bjorling@wdc.com> > Cc: Keith Busch <kbusch@kernel.org>; Damien Le Moal > <Damien.LeMoal@wdc.com>; Fam Zheng <fam@euphon.net>; Kevin Wolf > <kwolf@redhat.com>; qemu-block@nongnu.org; Niklas Cassel > <Niklas.Cassel@wdc.com>; Klaus Jensen <k.jensen@samsung.com>; qemu- > devel@nongnu.org; Alistair Francis <Alistair.Francis@wdc.com>; Philippe > Mathieu-Daudé <philmd@redhat.com> > Subject: Re: [PATCH v4 00/14] hw/block/nvme: Support Namespace Types and > Zoned Namespace Command Set > > On Sep 29 18:17, Matias Bjorling wrote: > > > > > > > -----Original Message----- > > > From: Klaus Jensen <its@irrelevant.dk> > > > Sent: Tuesday, 29 September 2020 20.00 > > > To: Keith Busch <kbusch@kernel.org> > > > Cc: Damien Le Moal <Damien.LeMoal@wdc.com>; Fam Zheng > > > <fam@euphon.net>; Kevin Wolf <kwolf@redhat.com>; qemu- > > > block@nongnu.org; Niklas Cassel <Niklas.Cassel@wdc.com>; Klaus > > > Jensen <k.jensen@samsung.com>; qemu-devel@nongnu.org; Alistair > > > Francis <Alistair.Francis@wdc.com>; Philippe Mathieu-Daudé > > > <philmd@redhat.com>; Matias Bjorling <Matias.Bjorling@wdc.com> > > > Subject: Re: [PATCH v4 00/14] hw/block/nvme: Support Namespace Types > > > and Zoned Namespace Command Set > > > > > > On Sep 29 10:29, Keith Busch wrote: > > > > On Tue, Sep 29, 2020 at 12:46:33PM +0200, Klaus Jensen wrote: > > > > > It is unmistakably clear that you are invalidating my arguments > > > > > about portability and endianness issues by suggesting that we > > > > > just remove persistent state and deal with it later, but > > > > > persistence is the killer feature that sets the QEMU emulated > > > > > device apart from other emulation options. It is not about using > > > > > emulation in production (because yeah, why would you?), but > > > > > persistence is what makes it possible to develop and test "zoned > > > > > FTLs" or something that > > > requires recovery at power up. > > > > > This is what allows testing of how your host software deals with > > > > > opened zones being transitioned to FULL on power up and the > > > > > persistent tracking of LBA allocation (in my series) can be used > > > > > to properly test error recovery if you lost state in the app. > > > > > > > > Hold up -- why does an OPEN zone transition to FULL on power up? > > > > The spec suggests it should be CLOSED. The spec does appear to > > > > support going to FULL on a NVM Subsystem Reset, though. Actually, > > > > now that I'm looking at this part of the spec, these implicit > > > > transitions seem a bit less clear than I expected. I'm not sure > > > > it's clear enough to evaluate qemu's compliance right now. > > > > > > > > But I don't see what testing these transitions has to do with > > > > having a persistent state. You can reboot your VM without tearing > > > > down the running QEMU instance. You can also unbind the driver or > > > > shutdown the controller within the running operating system. That > > > > should make those implicit state transitions reachable in order to > > > > exercise your FTL's recovery. > > > > > > > > > > Oh dear - don't "spec" with me ;) > > > > > > NVMe v1.4 Section 7.3.1: > > > > > > An NVM Subsystem Reset is initiated when: > > > * Main power is applied to the NVM subsystem; > > > * A value of 4E564D64h ("NVMe") is written to the NSSR.NSSRC > > > field; > > > * Requested using a method defined in the NVMe Management > > > Interface specification; or > > > * A vendor specific event occurs. > > > > > > In the context of QEMU, "Main power" is tearing down QEMU and > > > starting it from scratch. Just like on a "real" host, unbinding the > > > driver, rebooting or shutting down the controller does not cause a > > > subsystem reset (and does not cause the zones to change state). And > > > since the device does not indicate support for the optional > > > NSSR.NSSRC register, that way to initiate a subsystem cannot be used. > > > > > > The reason for moving to FULL is that write pointer updates are not > > > persisted on each advancement, only when the zone state changes. So > > > zones that were opened might have valid data, but invalid write pointer. > > > So the device transitions them to FULL as it is allowed to. > > > > > > > How about when one must also recover from intermediate states (i.e., > > open/closed upon power loss). For example, I don't hope a real SSD > > implementation transition zones to full when it has thousands of open > > simultaneously. That could be a disaster for the PE cycles, and a lot > > of media going to waste. One would want applications to support that > > kind of failure mode as well. > > Christ. The WDC Strike Force is really jumping out of lightspeed here. > I'm afraid I don't have an opposing force to engage with. So I'll be your only > boxing bag for the evening. > > As Keith just said, "Opened" is not a valid intial state. Didn't you write the > spec? ;) As for Closed, they will be brought up as is. Upon power failure, a zone in the Explicitly Opened state or the Implicitly Opened state, and has LBAs written, can either be transitioned to Full or Closed state by the controller. In the previous mail, I wanted to point out that if the intention of qemu was to test applications upon power failures, it could be beneficial to have an option that allowed transitioning open zones to closed upon power failure. Then applications can be tested with that in mind as well, without having access to an SSD that provided that kind of implementation. > > With that in mind, I'm not sure what you specifically refer to? I'll gently remind > you that the QEMU nvme device is not a real SSD and does not deal with NAND > so it does not really do any "recovering" of intermediate states on power on if > that is what you refer to?
> -----Original Message----- > From: Keith Busch <kbusch@kernel.org> > Sent: Tuesday, September 29, 2020 3:22 PM > To: Klaus Jensen <its@irrelevant.dk> > Cc: Dmitry Fomichev <Dmitry.Fomichev@wdc.com>; Kevin Wolf > <kwolf@redhat.com>; Fam Zheng <fam@euphon.net>; Damien Le Moal > <Damien.LeMoal@wdc.com>; qemu-block@nongnu.org; Niklas Cassel > <Niklas.Cassel@wdc.com>; Klaus Jensen <k.jensen@samsung.com>; qemu- > devel@nongnu.org; Alistair Francis <Alistair.Francis@wdc.com>; Philippe > Mathieu-Daudé <philmd@redhat.com>; Matias Bjorling > <Matias.Bjorling@wdc.com> > Subject: Re: [PATCH v4 00/14] hw/block/nvme: Support Namespace Types > and Zoned Namespace Command Set > > All, > > Let's de-escalate this, please. There's no reason to doubt Klaus wants > to see this to work well, just as everyone else does. We unfortunately > have conflicting proposals posted, and everyone is passionate enough > about their work, but please simmer down. > > As I mentioned earlier, I'd like to refocus on the basic implementation > and save the persistent state discussion once the core is solid. After > going through it all, I feel there's enough to discuss there to keep us > busy for little while longer. Additional comments on the code will be > coming from me later today. OK, I agree with this and I will not be replying to the email prior to this one it the thread. Let's calm down so we will be able to have a beer at a conference one day :) The only one thing that I would like to cover is lack of response to Klaus' ZNS patchset. Klaus, you are right to complain about it. Since discovering about the large backlog of NVMe patches that you had pending (something that we were not aware at the time of publishing our patches), we made the decision to rebase our series on top of the patches that you had posted before the publication time of WDC ZNS patchset. Since then, I got caught in the constant cycle of rebasing our patches on top of your series and that prevented me from doing much in terms of reviewing of your commits. Now, once we seem to catch up with the current head of development, I should be able to do more of this. There is absolutely no ill will involved :) Dmitry
v3 -> v4 - Fix bugs introduced in v2/v3 for QD > 1 operation. Now, all writes to a zone happen at the new write pointer variable, zone->w_ptr, that is advanced right after submitting the backend i/o. The existing zone->d.wp variable is updated upon the successful write completion and it is used for zone reporting. Some code has been split from nvme_finalize_zoned_write() function to a new function, nvme_advance_zone_wp(). - Make the code compile under mingw. Switch to using QEMU API for mmap/msync, i.e. memory_region...(). Since mmap is not available in mingw (even though there is mman-win32 library available on Github), conditional compilation is added around these calls to avoid undefined symbols under mingw. A better fix would be to add stub functions to softmmu/memory.c for the case when CONFIG_POSIX is not defined, but such change is beyond the scope of this patchset and it can be made in a separate patch. - Correct permission mask used to open zone metadata file. - Fold "Define 64 bit cqe.result" patch into ZNS commit. - Use clz64/clz32 instead of defining nvme_ilog2() function. - Simplify rpt_empty_id_struct() code, move nvme_fill_data() back to ZNS patch. - Fix a power-on processing bug. - Rename NVME_CMD_ZONE_APND to NVME_CMD_ZONE_APPEND. - Make the list of review comments addressed in v2 of the series (see below). v2 -> v3: - Moved nvme_fill_data() function to the NSTypes patch as it is now used there to output empty namespace identify structs. - Fixed typo in Maxim's email address. v1 -> v2: - Rebased on top of qemu-nvme/next branch. - Incorporated feedback from Klaus and Alistair. * Allow a subset of CSE log to be read, not the entire log * Assign admin command entries in CSE log to ACS fields * Set LPA bit 1 to indicate support of CSE log page * Rename CC.CSS value CSS_ALL_NSTYPES (110b) to CSS_CSI * Move the code to assign lbaf.ds to a separate patch * Remove the change in firmware revision * Change "driver" to "device" in comments and annotations * Rename ZAMDS to ZASL * Correct a few format expressions and some wording in trace event definitions * Remove validation code to return NVME_CAP_EXCEEDED error * Make ZASL to be equal to MDTS if "zone_append_size_limit" module parameter is not set * Clean up nvme_zoned_init_ctrl() to make size calculations less confusing * Avoid changing module parameters, use separate n/s variables if additional calculations are necessary to convert parameters to running values * Use NVME_DEFAULT_ZONE_SIZE to assign the default zone size value * Use default 0 for zone capacity meaning that zone capacity will be equal to zone size by default * Issue warnings if user MAR/MOR values are too large and have to be adjusted * Use unsigned values for MAR/MOR - Dropped "Simulate Zone Active excursions" patch. Excursion behavior may depend on the internal controller architecture and therefore be vendor-specific. - Dropped support for Zone Attributes and zoned AENs for now. These features can be added in a future series. - NS Types support is extended to handle active/inactive namespaces. - Update the write pointer after backing storage I/O completion, not before. This makes the emulation to run correctly in case of backing device failures. - Avoid division in the I/O path if the device zone size is a power of two (the most common case). Zone index then can be calculated by using bit shift. - A few reported bugs have been fixed. - Indentation in function definitions has been changed to make it the same as the rest of the code. Zoned Namespace (ZNS) Command Set is a newly introduced command set published by the NVM Express, Inc. organization as TP 4053. The main design goals of ZNS are to provide hardware designers the means to reduce NVMe controller complexity and to allow achieving a better I/O latency and throughput. SSDs that implement this interface are commonly known as ZNS SSDs. This command set is implementing a zoned storage model, similarly to ZAC/ZBC. As such, there is already support in Linux, allowing one to perform the majority of tasks needed for managing ZNS SSDs. The Zoned Namespace Command Set relies on another TP, known as Namespace Types (NVMe TP 4056), which introduces support for having multiple command sets per namespace. Both ZNS and Namespace Types specifications can be downloaded by visiting the following link - https://nvmexpress.org/wp-content/uploads/NVM-Express-1.4-Ratified-TPs.zip This patch series adds Namespace Types support and zoned namespace emulation capability to the existing NVMe PCI device. The patchset is organized as follows - The first several patches are preparatory and are added to allow for an easier review of the subsequent commits. The group of patches that follows adds NS Types support with only NVM Command Set being available. Finally, the last group of commits makes definitions and adds new code to support Zoned Namespace Command Set. Based-on: Message-ID: <20200729220638.344477-17-its@irrelevant.dk> Dmitry Fomichev (11): hw/block/nvme: Report actual LBA data shift in LBAF hw/block/nvme: Add Commands Supported and Effects log hw/block/nvme: Define trace events related to NS Types hw/block/nvme: Make Zoned NS Command Set definitions hw/block/nvme: Define Zoned NS Command Set trace events hw/block/nvme: Support Zoned Namespace Command Set hw/block/nvme: Introduce max active and open zone limits hw/block/nvme: Support Zone Descriptor Extensions hw/block/nvme: Add injection of Offline/Read-Only zones hw/block/nvme: Use zone metadata file for persistence hw/block/nvme: Document zoned parameters in usage text Niklas Cassel (3): hw/block/nvme: Introduce the Namespace Types definitions hw/block/nvme: Add support for Namespace Types hw/block/nvme: Add support for active/inactive namespaces block/nvme.c | 2 +- hw/block/nvme.c | 1987 +++++++++++++++++++++++++++++++++++++++-- hw/block/nvme.h | 180 ++++ hw/block/trace-events | 39 + include/block/nvme.h | 210 ++++- 5 files changed, 2351 insertions(+), 67 deletions(-)