[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
RE: [PATCH 0/5] ARM virt: Add NVDIMM support
|
From: |
Shameerali Kolothum Thodi |
|
Subject: |
RE: [PATCH 0/5] ARM virt: Add NVDIMM support |
|
Date: |
Mon, 25 Nov 2019 16:25:54 +0000 |
Hi Igor,
> -----Original Message-----
> From: Igor Mammedov [mailto:address@hidden]
> Sent: 25 November 2019 15:46
> To: Shameerali Kolothum Thodi <address@hidden>
> Cc: Auger Eric <address@hidden>; address@hidden;
> address@hidden; address@hidden;
> address@hidden; xuwei (O) <address@hidden>;
> address@hidden; Linuxarm <address@hidden>
> Subject: Re: [PATCH 0/5] ARM virt: Add NVDIMM support
>
> On Mon, 25 Nov 2019 13:20:02 +0000
> Shameerali Kolothum Thodi <address@hidden> wrote:
>
> > Hi Eric/Igor,
> >
> > > -----Original Message-----
> > > From: Shameerali Kolothum Thodi
> > > Sent: 22 October 2019 15:05
> > > To: 'Auger Eric' <address@hidden>; address@hidden;
> > > address@hidden; address@hidden
> > > Cc: address@hidden; address@hidden; xuwei (O)
> > > <address@hidden>; address@hidden; Linuxarm
> > > <address@hidden>
> > > Subject: RE: [PATCH 0/5] ARM virt: Add NVDIMM support
>
> not related to problem discussed in this patch but you probably
> need to update docs/specs/acpi_nvdimm.txt to account for your changes
Ok.
> > >
> >
> > [..]
> >
> > > > one question: I noticed that when a NVDIMM slot is hotplugged one get
> > > > the following trace on guest:
> > > >
> > > > nfit ACPI0012:00: found a zero length table '0' parsing nfit
> > > > pmem0: detected capacity change from 0 to 1073741824
> > > >
> > > > Have you experienced the 0 length trace?
> > >
> > > I double checked and yes that trace is there. And I did a quick check with
> > > x86 and it is not there.
> > >
> > > The reason looks like, ARM64 kernel receives an additional 8 bytes size
> when
> > > the kernel evaluates the "_FIT" object.
> > >
> > > For the same test scenario, Qemu reports a FIT buffer size of 0xb8 and
> > >
> > > X86 Guest kernel,
> > > [ 1.601077] acpi_nfit_init: data 0xffff8a273dc12b18 sz 0xb8
> > >
> > > ARM64 Guest,
> > > [ 0.933133] acpi_nfit_init: data 0xffff00003cbe6018 sz 0xc0
> > >
> > > I am not sure how that size gets changed for ARM which results in
> > > the above mentioned 0 length trace. I need to debug this further.
> > >
> > > Please let me know if you have any pointers...
> >
> > I spend some time debugging this further and it looks like the AML code
> > behaves differently on x86 and ARM64.
> FIT table is built dynamically and you are the first to debug
> such issue.
> (apart from the author the NVDIMM code.
:)
> btw: why NVDIMM author is not on CC list???)
Right. Missed that. CCd.
>
> > Booted guest with nvdimm mem, and used SSDT override with dbg prints
> > added,
> >
> > -object memory-backend-ram,id=mem1,size=1G \
> > -device nvdimm,id=dimm1,memdev=mem1 \
> >
> > On X86,
> > -----------
> >
> > [Qemu]NVDIMM:nvdimm_dsm_func_read_fit: Read FIT: offset 0 FIT size 0xb8
> Dirty Yes.
> > [Qemu]NVDIMM:nvdimm_dsm_func_read_fit: read_fit_out buf size 0xc0
> func_ret_status 0
> >
> > [AML]"NVDIMM-NCAL: Rcvd RLEN 000000C0"
> > [AML]"NVDIMM-NCAL: Creating OBUF with 000005E0 bits"
> > [AML]"NVDIMM-NCAL: Created BUF(Local7) size 000000BC"
> > [AML]"NVDIMM-RFIT Rcvd buf size 000000BC"
> > [AML]"NVDIMM-RFIT Created NVDR.RFIT.BUFF size 000000B8"
> > [AML]"NVDIMM-FIT: Rcvd buf size 000000B8"
> >
> > [Qemu]NVDIMM:nvdimm_dsm_func_read_fit: Read FIT: offset 0xb8 FIT size
> 0xb8 Dirty No.
> > [Qemu]NVDIMM:nvdimm_dsm_func_read_fit: read_fit_out buf size 0x8
> func_ret_status 0
> >
> > [AML]"NVDIMM-NCAL: Rcvd RLEN 00000008"
> > [AML]"NVDIMM-NCAL: Creating OBUF with 00000020 bits"
> > [AML]"NVDIMM-NCAL: Created BUF(Local7) size 00000004"
> > [AML]"NVDIMM-RFIT Rcvd buf size 00000004"
> > [AML]"NVDIMM-FIT: Rcvd buf size 00000000"
> > [AML]"NVDIMM-FIT: _FIT returned size 000000B8"
> >
> > [ KERNEL] acpi_nfit_init: NVDIMM: data 0xffff9855bb9a7518 sz 0xb8 -->
> Guest receives correct size(0xb8) here
> >
> > On ARM64,
> > ---------------
> >
> > [Qemu]NVDIMM:nvdimm_dsm_func_read_fit: Read FIT: offset 0 FIT size 0xb8
> Dirty Yes.
> > [Qemu]VDIMM:nvdimm_dsm_func_read_fit: read_fit_out buf size 0xc0
> func_ret_status 0
> >
> > [AML]"NVDIMM-NCAL: Rcvd RLEN 00000000000000C0"
> > [AML]"NVDIMM-NCAL: Creating OBUF with 00000000000005E0 bits"
> > [AML]"NVDIMM-NCAL: Created BUF(Local7) size 00000000000000BC"
> > [AML]"NVDIMM-RFIT Rcvd buf size 00000000000000BC"
> > [AML]"NVDIMM-RFIT Created NVDR.RFIT.BUFF size 00000000000000B8"
> > [AML]"NVDIMM-FIT: Rcvd buf size 00000000000000B8"
> >
> > [Qemu]NVDIMM:nvdimm_dsm_func_read_fit: Read FIT: offset 0xb8 FIT size
> 0xb8 Dirty No.
> > [Qemu]NVDIMM:nvdimm_dsm_func_read_fit: read_fit_out buf size 0x8
> func_ret_status 0
> >
> > [AML]"NVDIMM-NCAL: Rcvd RLEN 0000000000000008"
> > [AML]"NVDIMM-NCAL: Creating OBUF with 0000000000000020 bits" --> All
> looks same as x86 up to here.
> > [AML]"NVDIMM-NCAL: Created BUF(Local7) size 0000000000000008" --->
> The size goes wrong. 8 bytes instead of 4!.
>
> > [AML]"NVDIMM-RFIT Rcvd buf size 0000000000000008"
> > [AML]"NVDIMM-RFIT Created NVDR.RFIT.BUFF size 0000000000000004"
> > [AML]"NVDIMM-FIT: Rcvd buf size 0000000000000008" --> Again size goes
> wrong. 8 bytes instead of 4!.
>
> Local1 = SizeOf (Local0)
> printf("NVDIMM-RFIT Rcvd buf size %o", Local1)
>
> Local1 -= 0x04
> ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ here you get -4 so sizes you
> see are consistent
Hmm...IIUC, Method(RFIT, ) creates a buffer of size 0x4 as per
this,
"NVDIMM-RFIT Created NVDR.RFIT.BUFF size 0000000000000004"
And this is the buffer received by Method(_FIT, ),
Local0 = RFIT (Local3)
Local1 = SizeOf (Local0)
printf("NVDIMM-FIT: Rcvd buf size %o", Local1)
But "NVDIMM-FIT: Rcvd buf size 0000000000000008".
As you can see above, in the first iteration the buffer size received by _FIT is
actually what RFIT created(0xB8).
>
> If ((Local1 == Zero))
> {
> Return (Buffer (Zero){})
> }
>
> CreateField (Local0, 0x20, (Local1 << 0x03), BUFF)
> printf("NVDIMM-RFIT Created NVDR.RFIT.BUFF size %o",
> Local1)
>
>
> > [Qemu]NVDIMM:nvdimm_dsm_func_read_fit: Read FIT: offset 0xc0 FIT size
> 0xb8 Dirty No. -->Another read is attempted
> > [Qemu]NVDIMM:nvdimm_dsm_func_read_fit: read_fit_out buf size 0x8
> func_ret_status 3 --> Error status returned
>
> status 3 means that QEMU didn't like content of NRAM, and there is only
> 1 place like this in nvdimm_dsm_func_read_fit()
> if (read_fit->offset > fit->len) {
> func_ret_status = NVDIMM_DSM_RET_STATUS_INVALID;
> goto exit;
> }
>
> so I'd start looking from here and check that QEMU gets expected data
> in nvdimm_dsm_write(). In other words I'd try to trace/compare
> content of DSM buffer (from qemu side).
I had printed the DSM buffer previously and it looked same, I will double check
that.
>
> > [AML]"NVDIMM-NCAL: Rcvd RLEN 0000000000000008"
> > [AML]"NVDIMM-NCAL: Creating OBUF with 0000000000000020 bits"
> > [AML]"NVDIMM-NCAL: Created BUF(Local7) size 0000000000000008"
> > [AML]"NVDIMM-FIT: Rcvd buf size 0000000000000000"
> RFIT returns 0-sized buffer in case of error
Yes.
If ((Zero != STAU))
{
Return (Buffer (Zero){})
}
> > [AML]"NVDIMM-FIT: _FIT returned size 00000000000000C0" --> Wrong
> size returned.
> after that it goes
>
> Local0 = Buffer (Zero){}
> Local0 = RFIT (Local3)
> Local1 = SizeOf (Local0)
> printf("NVDIMM-FIT: Rcvd buf size %o", Local1)
> If ((RSTA == 0x0100))
> {
> Local2 = Buffer (Zero){}
> Local3 = Zero
> }
> Else
> {
> If ((Local1 == Zero))
> --> here, probably on the second iteration of the loop taking Zero
> size as EOF sign but without checking for any error (RSTA !=0)
> and returning partial buffer only
I guess you meant the third iteration. Yes, it doesn't check for RSTA != 0
case. And in the second iteration, since status == 0 and buffer len received is
8 bytes,
Concatenate (Local2, Local0, Local2)
The final buffer returned to Guest becomes 0xC0 (0xb8 + 0x8).
> {
> printf("NVDIMM-FIT: _FIT returned size %o",
> SizeOf(Local2))
> Return (Local2)
> }
>
> relevant code in QEMU that builds this AML is in nvdimm_build_fit()
>
> > [ KERNEL] acpi_nfit_init: NVDIMM: data 0xffff0000fc57ce18 sz 0xc0
> -->Kernel gets 0xc0 instead of 0xb8
> >
> >
> > It looks like the aml, "CreateField (ODAT, Zero, Local1, OBUF)" goes wrong
> > for
> > ARM64 when the buffer is all zeroes. My knowledge on aml is very limited and
> not
> > sure this is a 32/64bit issue or not. I am attaching the SSDT files with the
> above
> > dbg prints added. Could you please take a look and let me know what actually
> is
> > going on here...
>
> diff -u SSDT-dbg-x86.dsl SSDT-dbg-arm64.dsl
> --- SSDT-dbg-x86.dsl 2019-11-25 14:50:22.024983026 +0100
> +++ SSDT-dbg-arm64.dsl 2019-11-25 14:50:06.740690645 +0100
> @@[...]
> @@ -28,7 +28,7 @@
> Method (NCAL, 5, Serialized)
> {
> Local6 = MEMA /* \MEMA */
> - OperationRegion (NPIO, SystemIO, 0x0A18, 0x04)
> + OperationRegion (NPIO, SystemMemory, 0x09090000,
> 0x04)
> that's fine and matches your code
>
> OperationRegion (NRAM, SystemMemory, Local6, 0x1000)
> Field (NPIO, DWordAcc, NoLock, Preserve)
> {
> @@ -210,6 +210,6 @@
> }
> }
>
> - Name (MEMA, 0xBFBFD000)
> + Name (MEMA, 0xFFFF0000)
>
> However value here is suspicious. If I recall right it should
> point to DMS buffer allocated by firmware somewhere in the guest RAM.
But then it will be horribly wrong and will result in inconsistent behavior
as well, right?
Thanks,
Shameer
>
> }
>
> > Much appreciated,
> > Shameer.
> >
> >