Re: how to disable(not ignore) disk discard feature

[Frantisek Rysanek]

I've taken a look at the qemu source code (7.2.0 happens to be 
unpacked on my hard drive).

I have noticed that there probably two separate implementations of 
virtual disk devices supporting  unmap/discard.
My scrapbook notes on this chapter:


1)   hw/scsi/scsi-disk.c:

// The following reflects in max_unmap_size and max_unmap_sectors
// in VPD page 0xb0 AKA "block limits". 
// How 'bout if we redefine this to 0 ?
#define DEFAULT_MAX_UNMAP_SIZE      (1 * GiB)

Curiously, the Logical Block Provisioning VPD page (0x82) is not 
considered...
in function scsi_disk_emulate_vpd_page()



2) the virtio collective:

hw/block/virtio-blk.c
hw/block/virtio-blk-common.c
include/standard-headers/linux/virtio_blk.h
VIRTIO_BLK_F_DISCARD
conf.max_discard_sectors
block/export/vhost-user-blk-server.c
VIRTIO_BLK_MAX_DISCARD_SECTORS
struct virtio_blk_config
virtio_blk_set_config
virtio_blk_handle_scsi_req() is a passthrough to async IO ?

Here, it appears that VIRTIO_BLK_F_DISCARD is a feature of the "disk 
controller" (virtio itself) i.e. it is global, but there's a 
device-private member "max_discard_sectors" which sounds interesting, 
perhaps could be set to 0.
Somewhere in the code, I've spotted a check, if VIRTIO_BLK_F_DISCARD 
is wholesale enabled, that max_discard_sectors must be non-zero, 
otherwise virtio-blk throws a fit (error).

Curiously the virtio_blk_set_config() appears to have a single 
settable option, which is "write cache enable" (or wce for short).
Perhaps this could be extended to have the max_discard_sectors 
affected as well.

But, I'm also wondering how the guest-side virtio drivers go about 
inquiring the disk devices. If they tap directly into the 
virtio_blk_config and maybe even the virtio global config, then it 
may as well be hard cheese :-) I.e. the guest OS will get informed 
that the virtio-blk devices *do* categorically support Discard, and 
may throw erros at the guest side, if it turns out that the maximum 
Discard length is 0 sectors :-)
The truth is, that I haven't found the exact mechanism, how the 
guest-side drivers ask the host-side Virtio stack / how it inquires 
the virtio devices.
And, there is also a "SCSI command handler" within 
hw/block/virtio-blk.c
which seems to just take any incoming SCSI commands and shovel them 
"down some further pipeline" - looks like an async IO ioctl() or 
something... makes me wonder if "there someplace else" the Inquiry 
could actually be handled, including the VPD page 0x82, which we 
could catch and maybe rig...

I just don't know. This is over my head. I'm done :-)

Frank


BTW see also

SCSI Commands Reference Manual
https://www.seagate.com/files/staticfiles/support/docs/manual/Interfac
e%20manuals/100293068j.pdf

5.4.13 Logical Block Provisioning VPD page (82h)
That's where the LBPU flag is defined...