Hi Simon, I'd like to resume this discussion.
On Thu, Oct 14, 2021 at 02:55:36PM -0600, Simon Glass wrote: > Hi Takahiro, > > On Thu, 14 Oct 2021 at 02:03, AKASHI Takahiro > <takahiro.aka...@linaro.org> wrote: > > > > Simon, > > > > On Wed, Oct 13, 2021 at 12:05:58PM -0600, Simon Glass wrote: > > > Hi Takahiro, > > > > > > On Tue, 12 Oct 2021 at 19:32, AKASHI Takahiro > > > <takahiro.aka...@linaro.org> wrote: > > > > > > > > On Tue, Oct 12, 2021 at 11:14:17AM -0400, Tom Rini wrote: > > > > > On Mon, Oct 11, 2021 at 10:14:00AM -0600, Simon Glass wrote: > > > > > > Hi Heinrich, > > > > > > > > > > > > On Mon, 11 Oct 2021 at 09:02, Heinrich Schuchardt > > > > > > <xypron.g...@gmx.de> wrote: > > > > > > > > > > > > > > > > > > > > > > > > > > > > On 10/11/21 16:54, Simon Glass wrote: > > > > > > > > Hi Takahiro, > > > > > > > > > > > > > > > > On Sun, 10 Oct 2021 at 20:29, AKASHI Takahiro > > > > > > > > <takahiro.aka...@linaro.org> wrote: > > > > > > > >> > > > > > > > >> Heinrich, > > > > > > > >> > > > > > > > >> On Fri, Oct 08, 2021 at 10:23:52AM +0200, Heinrich Schuchardt > > > > > > > >> wrote: > > > > > > > >>> > > > > > > > >>> > > > > > > > >>> On 10/8/21 02:51, AKASHI Takahiro wrote: > > > > > > > >>>> On Mon, Oct 04, 2021 at 12:27:59PM +0900, AKASHI Takahiro > > > > > > > >>>> wrote: > > > > > > > >>>>> On Fri, Oct 01, 2021 at 11:30:37AM +0200, Heinrich > > > > > > > >>>>> Schuchardt wrote: > > > > > > > >>>>>> > > > > > > > >>>>>> > > > > > > > >>>>>> On 10/1/21 07:01, AKASHI Takahiro wrote: > > > > > > > >>>>>>> UCLASS_PARTITION device will be created as a child node of > > > > > > > >>>>>>> UCLASS_BLK device. > > > > > > > >>>>>>> > > > > > > > >>>>>>> Signed-off-by: AKASHI Takahiro > > > > > > > >>>>>>> <takahiro.aka...@linaro.org> > > > > > > > >>>>>>> --- > > > > > > > >>>>>>> drivers/block/blk-uclass.c | 111 > > > > > > > >>>>>>> +++++++++++++++++++++++++++++++++++++ > > > > > > > >>>>>>> include/blk.h | 9 +++ > > > > > > > >>>>>>> include/dm/uclass-id.h | 1 + > > > > > > > >>>>>>> 3 files changed, 121 insertions(+) > > > > > > > >>>>>>> > > > > > > > >>>>>>> diff --git a/drivers/block/blk-uclass.c > > > > > > > >>>>>>> b/drivers/block/blk-uclass.c > > > > > > > >>>>>>> index 83682dcc181a..dd7f3c0fe31e 100644 > > > > > > > >>>>>>> --- a/drivers/block/blk-uclass.c > > > > > > > >>>>>>> +++ b/drivers/block/blk-uclass.c > > > > > > > >>>>>>> @@ -12,6 +12,7 @@ > > > > > > > >>>>>>> #include <log.h> > > > > > > > >>>>>>> #include <malloc.h> > > > > > > > >>>>>>> #include <part.h> > > > > > > > >>>>>>> +#include <string.h> > > > > > > > >>>>>>> #include <dm/device-internal.h> > > > > > > > >>>>>>> #include <dm/lists.h> > > > > > > > >>>>>>> #include <dm/uclass-internal.h> > > > > > > > >>>>>>> @@ -695,6 +696,44 @@ int blk_unbind_all(int if_type) > > > > > > > >>>>>>> return 0; > > > > > > > >>>>>>> } > > > > > > > >>>>>>> > > > > > > > >>>>>>> +int blk_create_partitions(struct udevice *parent) > > > > > > > >>>>>>> +{ > > > > > > > >>>>>>> + int part, count; > > > > > > > >>>>>>> + struct blk_desc *desc = dev_get_uclass_plat(parent); > > > > > > > >>>>>>> + struct disk_partition info; > > > > > > > >>>>>>> + struct disk_part *part_data; > > > > > > > >>>>>>> + char devname[32]; > > > > > > > >>>>>>> + struct udevice *dev; > > > > > > > >>>>>>> + int ret; > > > > > > > >>>>>>> + > > > > > > > >>>>>>> + if (!CONFIG_IS_ENABLED(PARTITIONS) || > > > > > > > >>>>>>> + !CONFIG_IS_ENABLED(HAVE_BLOCK_DEVICE)) > > > > > > > >>>>>>> + return 0; > > > > > > > >>>>>>> + > > > > > > > >>>>>>> + /* Add devices for each partition */ > > > > > > > >>>>>>> + for (count = 0, part = 1; part <= > > > > > > > >>>>>>> MAX_SEARCH_PARTITIONS; part++) { > > > > > > > >>>>>>> + if (part_get_info(desc, part, &info)) > > > > > > > >>>>>>> + continue; > > > > > > > >>>>>>> + snprintf(devname, sizeof(devname), "%s:%d", > > > > > > > >>>>>>> parent->name, > > > > > > > >>>>>>> + part); > > > > > > > >>>>>>> + > > > > > > > >>>>>>> + ret = device_bind_driver(parent, > > > > > > > >>>>>>> "blk_partition", > > > > > > > >>>>>>> + strdup(devname), > > > > > > > >>>>>>> &dev); > > > > > > > >>>>>>> + if (ret) > > > > > > > >>>>>>> + return ret; > > > > > > > >>>>>>> + > > > > > > > >>>>>>> + part_data = dev_get_uclass_plat(dev); > > > > > > > >>>>>>> + part_data->partnum = part; > > > > > > > >>>>>>> + part_data->gpt_part_info = info; > > > > > > > >>>>>>> + count++; > > > > > > > >>>>>>> + > > > > > > > >>>>>>> + device_probe(dev); > > > > > > > >>>>>>> + } > > > > > > > >>>>>>> + debug("%s: %d partitions found in %s\n", __func__, > > > > > > > >>>>>>> count, parent->name); > > > > > > > >>>>>>> + > > > > > > > >>>>>>> + return 0; > > > > > > > >>>>>>> +} > > > > > > > >>>>>>> + > > > > > > > >>>>>>> static int blk_post_probe(struct udevice *dev) > > > > > > > >>>>>>> { > > > > > > > >>>>>>> if (IS_ENABLED(CONFIG_PARTITIONS) && > > > > > > > >>>>>>> @@ -713,3 +752,75 @@ UCLASS_DRIVER(blk) = { > > > > > > > >>>>>>> .post_probe = blk_post_probe, > > > > > > > >>>>>>> .per_device_plat_auto = sizeof(struct blk_desc), > > > > > > > >>>>>>> }; > > > > > > > >>>>>>> + > > > > > > > >>>>>>> +static ulong blk_part_read(struct udevice *dev, lbaint_t > > > > > > > >>>>>>> start, > > > > > > > >>>>>>> + lbaint_t blkcnt, void *buffer) > > > > > > > >>>>>>> +{ > > > > > > > >>>>>>> + struct udevice *parent; > > > > > > > >>>>>>> + struct disk_part *part; > > > > > > > >>>>>>> + const struct blk_ops *ops; > > > > > > > >>>>>>> + > > > > > > > >>>>>>> + parent = dev_get_parent(dev); > > > > > > > >>>>>> > > > > > > > >>>>>> What device type will the parent have if it is a eMMC > > > > > > > >>>>>> hardware partition? > > > > > > > >>>>>> > > > > > > > >>>>>>> + ops = blk_get_ops(parent); > > > > > > > >>>>>>> + if (!ops->read) > > > > > > > >>>>>>> + return -ENOSYS; > > > > > > > >>>>>>> + > > > > > > > >>>>>>> + part = dev_get_uclass_plat(dev); > > > > > > > >>>>>> > > > > > > > >>>>>> You should check that we do not access the block device > > > > > > > >>>>>> past the > > > > > > > >>>>>> partition end: > > > > > > > >>>>> > > > > > > > >>>>> Yes, I will fix all of checks. > > > > > > > >>>>> > > > > > > > >>>>>> struct blk_desc *desc = dev_get_uclass_plat(parent); > > > > > > > >>>>>> if ((start + blkcnt) * desc->blksz < > > > > > > > >>>>>> part->gpt_part_info.blksz) > > > > > > > >>>>>> return -EFAULT. > > > > > > > >>>>>> > > > > > > > >>>>>>> + start += part->gpt_part_info.start; > > > > > > > >>>> > > > > > > > >>>> A better solution is: > > > > > > > >>>> if (start >= part->gpt_part_info.size) > > > > > > > >>>> return 0; > > > > > > > >>>> > > > > > > > >>>> if ((start + blkcnt) > part->gpt_part_info.size) > > > > > > > >>>> blkcnt = part->gpt_part_info.size - start; > > > > > > > >>>> start += part->gpt_part_info.start; > > > > > > > >>>> instead of returning -EFAULT. > > > > > > > >>>> (note that start and blkcnt are in "block".) > > > > > > > >>> > > > > > > > >>> What is your motivation to support an illegal access? > > > > > > > >>> > > > > > > > >>> We will implement the EFI_BLOCK_IO_PROTOCOL based on this > > > > > > > >>> function. The > > > > > > > >>> ReadBlocks() and WriteBlocks() services must return > > > > > > > >>> EFI_INVALID_PARAMETER if the read request contains LBAs that > > > > > > > >>> are not > > > > > > > >>> valid. > > > > > > > >> > > > > > > > >> I interpreted that 'LBA' was the third parameter to ReadBlocks > > > > > > > >> API, > > > > > > > >> and that if the starting block is out of partition region, we > > > > > > > >> should > > > > > > > >> return an error (and if not, we still want to trim IO request > > > > > > > >> to fit > > > > > > > >> into partition size as other OS's API like linux does). > > > > > > > >> Do you think it's incorrect? > > > > > > > > > > > > > > > > [..] > > > > > > > > > > > > > > > > Related to this patch I think that the partition type should be > > > > > > > > really > > > > > > > > be a child of the media device: > > > > > > > > > > > > > > > > - MMC > > > > > > > > |- BLK > > > > > > > > |- PARTITION > > > > > > > > |- BLK > > > > > > > > |- PARTITION > > > > > > > > |- BLK > > > > > > > > |- PARTITION > > > > > > > > |- BLK > > > > > > > > > > > > > > > > It seems more natural to me that putting the partitions under > > > > > > > > the > > > > > > > > top-level BLK device, so that BLK remains a 'terminal' device. > > > > > > > > > > > > > > > > The partition uclass is different from BLK, of course. It could > > > > > > > > contain information about the partition such as its partition > > > > > > > > number > > > > > > > > and UUID. > > > > > > > > > > > > > > Do you mean hardware partition here? Otherwise I would not know > > > > > > > what BLK > > > > > > > should model. > > > > > > > > > > > > I mean that (I think) we should not use BLK to model partitions. A > > > > > > BLK > > > > > > should just be a block device. > > > > > > > > > > > > I don't see any difference between a partition and a hardware > > > > > > partition. We presumably end up with a hierarchy though. Do we need > > > > > > a > > > > > > HWPARTITION uclass so we can handle the hardware partitions > > > > > > differently? > > > > > > > > > > Note that for eMMC devices, hardware partitions are different from > > > > > partition-table partitions. If you boot a system with an eMMC device > > > > > up > > > > > in Linux you typically get mmcblkN, mmcblkNboot0, mmcblkNboot1 and > > > > > mmcblkNrpmb, each of which are hardware partitions. It gets tricky in > > > > > U-Boot in that you can access each of these with 'mmc dev N M' where M > > > > > defaults to 0 and is the user partition (mmcblkN), 1/2 are boot0/boot1 > > > > > and 3 is the rpmb area. The 'mmc' command also allows, when possible > > > > > and implemented, configuring these partitions, again to the extent > > > > > allowed, documented and implemented. > > > > > > > > Thank you. That is exactly what I tried to mention in my reply > > > > at "part: call part_init() in blk_get_device_by_str() only for MMC" > > > > > > OK so it sounds like we agree that hwpartition and partition are > > > different things. > > > > Yes. > > Please note, IIUC, that > > * MMC hw partitions on a device are mapped to one udevice, differentiating > > them by blk_desc->hwpart. > > * Each NVME namespace on a device is mapped to a different udevice with > > a different blk_desc->devnum (and nvme_dev->ns_id). > > * Each UFS partition (or which is, I suppose, equivalent to scsi LUN) on > > a device is mapped to a different udevice with a different > > blk_desc->devnum > > (and blk_desc->lun). > > > > So even though those type of devices have some kind of hardware partitions, > > they are modelled differently in U-Boot. > > (Obviously, I might be wrong here as I'm not quite familiar yet.) > > > > > > > > > > ---8<--- > > > > # On the other hand, we have to explicitly switch "hw partitions" > > > > # with blk_select_hwpart_devnum() on MMC devices even though we use > > > > # the *same* udevice(blk_desc). > > > > --->8--- > > > > > > > > The problem with the current U-Boot driver model is that all of > > > > "mmcblkN, > > > > mmcblkNboot0, mmcblkNboot1 and mmcblkNrpmb" will be linked to the same > > > > udevice. We have to do "mmc dev N M" or call > > > > blk_select_hwpart[_devnum]() > > > > to distinguish them. > > > > > > Here's our chance to rethink this. What should the device hierarchy be > > > for an MMC device? I made a proposal further up the thread. > > > > Well, > > > > On Mon, Oct 11, 2021 at 11:41:02AM -0600, Simon Glass wrote: > > > On Mon, 11 Oct 2021 at 10:53, Heinrich Schuchardt <xypron.g...@gmx.de> > > > wrote: > > > > > > >>> [..] > > > > > > >>> Related to this patch I think that the partition type should be > > > > >>> really > > > > >>> be a child of the media device: > > > > >>> > > > > >>> - MMC > > > > >>> |- BLK > > > > >>> |- PARTITION > > > > >>> |- BLK > > > > >>> |- PARTITION > > > > >>> |- BLK > > > > >>> |- PARTITION > > > > >>> |- BLK > > > > >>> > > > > >>> It seems more natural to me that putting the partitions under the > > > > >>> top-level BLK device, so that BLK remains a 'terminal' device. > > > > >>> > > > > >>> The partition uclass is different from BLK, of course. It could > > > > >>> contain information about the partition such as its partition number > > > > >>> and UUID. > > > > Yeah, but there is always 1-to-1 mapping between a partition and > > a block (for a partition), so I still wonder whether it makes sense > > to model partitions in the way above. > > > > Alternatively, the following hierarchy also makes some sense. > > (This is not what I have in my RFC though.) > > - MMC > > |- BLK (whole disk with part=0) > > |- BLK (partition 1) > > |- BLK (partition 2) > > |- BLK (partition 3) > > > > or > > > > - MMC > > |- DISK (whole disk) > > ||- BLK (partition 0) > > ||- BLK (partition 1) > > ||- BLK (partition 2) > > ||- BLK (partition 3) > > > > Here > > MMC: provides read/write operations (via blk_ops) > > DISK: holds a geometry of a whole disk and other info > > BLK: partition info (+ blk_ops + geo) (part=0 means a while disk) > > Where does this leave hwpart? Are we giving up on that? No, not at all :) I'm thinking of dealing with hw partitions as independent BLK devices. This is already true for NVME (namespaces) and UFS (LUNs)(not sure, though). For MMC, struct blk_desc has 'hwpart' field to indicate a hw partition and Apparently, it will be easy to have different BLK devices with different hwpart's. (Then we will have to add a probe function for hw partitions.) > Both of these make some sense to me, although I'm not sure what the > second one buys us. Can you explain that? Is it to deal with hwpart? So, - MMC (bus controller) |- BLK (device/hw partition:user data) ||- DISK (partition 0 == a whole device) ||- DISK (partition 1) ||- DISK (partition 2) ||- DISK (partition 3) |- BLK (device/hw partition:boot0) ||- DISK (partition 0 == a whole device) |- BLK (device/hw partition:boot0) ||- DISK (partition 0 == a whole device) |- BLK (device/hw partition:rpmb) -- this is NOT a 'block' device, though. ||- DISK (partition 0 == a whole device) MMC: provides access methods (via blk_ops) BLK: represents a physical device and holds a geometry of the whole device and other info DISK: block-access entities with partition info (part=0 means a while disk) (MMC, BLK are of current implementation.) To avoid confusion, UCLASS_PARTITION is renamed to UCLASS_DISK with a little modified semantics. The name can be seen aligned with 'disk/' directory for sw partitions. Partition 0 expectedly behaves in the same way as an existing BLK. With this scheme, I assume that we should thoroughly use new interfaces dev_read(struct udevice *dev, lbaint_t start, lbaint_t blkcnt, void *buffer); dev_write(struct udevice *dev, lbaint_t start, lbaint_t blkcnt, void *buffer); for block-level operations with DISK devices. ^^^^ The legacy interfaces with blk_desc's in BLK devices: blk_dread(struct blk_desc *block_dev, lbaint_t start, lbaint_t blkcnt, void *buffer) blk_dwrite(struct blk_desc *block_dev, lbaint_t start, lbaint_t blkcnt, void *buffer)l are to be retained, at least, during the transition period (mostly for existing filesystems and commands). > The name 'disk' is pretty awful though, these days. Think so? Honestly, I'd like to rename BLK to DISK (or BLK_MEDIA) and rename DISK to BLK to reflect their rolls :) > If we want to iterate through all the partition tables across all > devices, we could do that with a partition uclass. We could support > different types of partition (s/w and h/w) with the same device > driver. > > I think conceptually it is cleaner to have a partition uclass but I do > agree that it corresponds 100% to BLK, so maybe there is little value > in practice. But which device holds the partition table in its > dev_get_priv()? Do you think that some device should have "partition table" info in its inner data structure of udevice? BLK-DISK relationship can represent a partition table in some way, and MMC-BLK can model hw partitioning. Thanks, -Takahiro Akashi > > > > > > >> Do you mean hardware partition here? Otherwise I would not know what > > > > >> BLK > > > > >> should model. > > > > > > > > > > I mean that (I think) we should not use BLK to model partitions. A BLK > > > > > should just be a block device. > > > > > > > > That is fine. But this implies that a software partition is the child of > > > > a block partition and not the other way round. So the tree should like: > > > > > > > > MMC > > > > |- BLK (user hardware partition) > > > > ||- PARTITION 1 (software partition) > > > > ||- PARTITION 2 (software partition) > > > > |... > > > > ||- PARTITION n (software partition) > > > > |- BLK (rpmb hardware partition) > > > > |- BLK (boot0 hardware partition) > > > > |- BLK (boot1 hardware partition) > > > > > > I presume you meant to include a BLK device under each PARTITION? > > > > > > But anyway, I was more thinking of this: > > > > > > MMC > > > | HWPARTITION rpmb > > > || BLK whole rpmb > > > || PARTITION 1 > > > ||| BLK > > > || PARTITION 2 > > > ||| BLK > > > || PARTITION 3 > > > > Do we have any reason to model a RPMB partition as a block device? > > For linux, at least, mmcblkrpmb looks to be a character device. > > > > > ||| BLK > > > | HWPARTITION boot0 > > > || BLK > > > (maybe have PARTITION in here too? > > > > I don't know how boot partitions are used on a production system. > > It's unlikely to have partitions on them given the purpose of "boot" > > partitions? > > That's true. So likely they will not be used. > > > > > > | HWPARTITION boot1 > > > (maybe have PARTITION in here too? > > > || BLK > > > > > > > > > > > > > > > > > I don't see any difference between a partition and a hardware > > > > > partition. We presumably end up with a hierarchy though. Do we need a > > > > > HWPARTITION uclass so we can handle the hardware partitions > > > > > differently? > > > > > > > > Software partitions are defined and discovered via partition tables. > > > > Hardware partitions are defined in a hardware specific way. > > > > > > > > All software partitions map to HD() device tree nodes in UEFI. > > > > An MMC device maps to an eMMC() node > > > > MMC hardware partitions are mapped to Ctrl() nodes by EDK II. We should > > > > do the same in U-Boot. > > > > An SD-card maps to an SD() node. > > > > An NVMe namespace maps to a NVMe() node. > > > > An SCSI LUN maps to a Scsi() node. > > > > SCSI channels of multiple channel controllers are mapped to Ctrl() > > > > nodes. > > > > > > I'm not quite sure about the terminology here. I'm not even talking > > > about UEFI, really, just how best to model this stuff in U-Boot. > > > > In UEFI world, each efi_disk has its own device path to identify the device. > > For example, here is a text representation of device path for a scsi disk > > partition: > > > > /VenHw(e61d73b9-a384-4acc-aeab-82e828f3628b)/Scsi(0,0)/HD(1,GPT,ce86c5a7-b32a-488f-a346-88fe698e0edc,0x22,0x4c2a) > > > > which is set to be created from a corresponding udevice (more strictly > > blkc_desc + part). > > > > So the issue Heinrich raised here is a matter of implementation of > > this conversion (software partitions, and SCSI channels?) as well as > > a modeling for some device type on U-Boot, i.e. MMC hardware partitions. > > Yes I see that. It's just that we should get our house in order first, > since these discussions didn't happen when the EFI layer was written 6 > years ago. If we have a good model for partitions (not just block > devices) in U-Boot then it should be easier to map EFI onto it. > > Regards, > Simon > > > > > > -Takahiro Akashi > > > > > In U-Boot, UCLASS_SCSI should be a SCSI controller, not a device, > > > right? I'm a little worried it is not modelled correctly. After all, > > > what is the parent of a SCSI device? > > > > > > > > > > > The simple file protocol is only provided by HD() nodes and not by nodes > > > > representing hardware partitions. If the whole hardware partition is > > > > formatted as a file system you would still create a HD() node with > > > > partition number 0. > > > > > > Regards, > > > Simon > > --- > > > > > > > > > > > When it comes to UEFI, I hope we can currently support hw partitions > > > > in this way: > > > > => efidebug add boot -b 1 FOO mmc 0.1 /foo.bin "" > > > > (".1" is a key, I have never tried this syntax though.) > > > > > > > > But probably its device path won't be properly formatted > > > > as expected as Heinrich suggested. > > > > > > > > -Takahiro Akashi > > > > > > > > > > > > > In terms of modeling, this is akin to how if you use a USB card reader > > > > > that supports 4 different form-factor cards, you can end up with 4 > > > > > different devices showing up in Linux (if you have one of the nice > > > > > card > > > > > readers that supports multiple cards at once). > > > > > > > > > > -- > > > > > Tom > > > > > > > > > > > > > > Regards, > > > Simon