Re: [PATCH v2 04/15] ACPI: table: replace __attribute__((packed)) by __packed

[Zhang Rui]

On Tue, 2021-03-30 at 08:14 +0000, David Laight wrote:
> From: Zhang Rui
> > Sent: 30 March 2021 09:00
> > To: Xiaofei Tan <tanxiaofei@xxxxxxxxxx>; David Laight <
> > David.Laight@xxxxxxxxxx>; rjw@xxxxxxxxxxxxx;
> > lenb@xxxxxxxxxx; bhelgaas@xxxxxxxxxx
> > Cc: linux-acpi@xxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx; 
> > linux-pci@xxxxxxxxxxxxxxx;
> > linuxarm@xxxxxxxxxxxxx
> > Subject: Re: [PATCH v2 04/15] ACPI: table: replace
> > __attribute__((packed)) by __packed
> > 
> > On Tue, 2021-03-30 at 15:31 +0800, Zhang Rui wrote:
> > > On Tue, 2021-03-30 at 10:23 +0800, Xiaofei Tan wrote:
> > > > Hi David,
> > > > 
> > > > On 2021/3/29 18:09, David Laight wrote:
> > > > > From: Xiaofei Tan
> > > > > > Sent: 27 March 2021 07:46
> > > > > > 
> > > > > > Replace __attribute__((packed)) by __packed following the
> > > > > > advice of checkpatch.pl.
> > > > > > 
> > > > > > Signed-off-by: Xiaofei Tan <tanxiaofei@xxxxxxxxxx>
> > > > > > ---
> > > > > >  drivers/acpi/acpi_fpdt.c | 6 +++---
> > > > > >  1 file changed, 3 insertions(+), 3 deletions(-)
> > > > > > 
> > > > > > diff --git a/drivers/acpi/acpi_fpdt.c
> > > > > > b/drivers/acpi/acpi_fpdt.c
> > > > > > index a89a806..690a88a 100644
> > > > > > --- a/drivers/acpi/acpi_fpdt.c
> > > > > > +++ b/drivers/acpi/acpi_fpdt.c
> > > > > > @@ -53,7 +53,7 @@ struct resume_performance_record {
> > > > > >  	u32 resume_count;
> > > > > >  	u64 resume_prev;
> > > > > >  	u64 resume_avg;
> > > > > > -} __attribute__((packed));
> > > > > > +} __packed;
> > > > > > 
> > > > > >  struct boot_performance_record {
> > > > > >  	struct fpdt_record_header header;
> > > > > > @@ -63,13 +63,13 @@ struct boot_performance_record {
> > > > > >  	u64 bootloader_launch;
> > > > > >  	u64 exitbootservice_start;
> > > > > >  	u64 exitbootservice_end;
> > > > > > -} __attribute__((packed));
> > > > > > +} __packed;
> > > > > > 
> > > > > >  struct suspend_performance_record {
> > > > > >  	struct fpdt_record_header header;
> > > > > >  	u64 suspend_start;
> > > > > >  	u64 suspend_end;
> > > > > > -} __attribute__((packed));
> > > > > > +} __packed;
> > > > > 
> > > > > My standard question about 'packed' is whether it is actually
> > > > > needed.
> > > > > It should only be used if the structures might be misaligned
> > > > > in
> > > > > memory.
> > > > > If the only problem is that a 64bit item needs to be 32bit
> > > > > aligned
> > > > > then a suitable type should be used for those specific
> > > > > fields.
> > > > > 
> > > > > Those all look very dubious - the standard header isn't
> > > > > packed
> > > > > so everything must eb assumed to be at least 32bit aligned.
> > > > > 
> > > > > There are also other sub-structures that contain 64bit
> > > > > values.
> > > > > These don't contain padding - but that requires 64bit
> > > > > alignement.
> > > > > 
> > > > > The only problematic structure is the last one - which would
> > > > > have
> > > > > a 32bit pad after the header.
> > > > > Is this even right given than there are explicit alignment
> > > > > pads
> > > > > in some of the other structures.
> > > > > 
> > > > > If 64bit alignment isn't guaranteed then a '64bit aligned to
> > > > > 32bit'
> > > > > type should be used for the u64 fields.
> > > > > 
> > > > 
> > > > Yes, some of them has been aligned already, then nothing
> > > > changed
> > > > when
> > > > add this "packed ". Maybe the purpose of the original author is
> > > > for
> > > > extension, and can tell others that this struct need be packed.
> > > > 
> > > 
> > > The patch is upstreamed recently but it was made long time ago.
> > > I think the original problem is that one of the address, probably
> > > the
> > > suspend_performance record, is not 64bit aligned, thus we can not
> > > read
> > > the proper content of suspend_start and suspend_end, mapped from
> > > physical memory.
> > > 
> > > I will try to find a machine to reproduce the problem with all
> > > __attribute__((packed)) removed to double confirm this.
> > > 
> > 
> > So here is the problem, without __attribute__((packed))
> > 
> > [    0.858442] suspend_record: 0xffffaad500175020
> > /sys/firmware/acpi/fpdt/suspend/suspend_end_ns:addr:
> > 0xffffaad500175030, 15998179292659843072
> > /sys/firmware/acpi/fpdt/suspend/suspend_start_ns:addr:
> > 0xffffaad500175028, 0
> > 
> > suspend_record is mapped to 0xffffaad500175020, and it is combined
> > with
> > one 32bit header and two 64bit fields (suspend_start and
> > suspend_end),
> > this is how it is located in physical memory.
> > So the addresses of the two 64bit fields are actually not 64bit
> > aligned.
> > 
> > David,
> > Is this the "a 64bit item needs to be 32bit aligned" problem you
> > referred?
> > If yes, what is the proper fix? should I used two 32bits for each
> > of
> > the field instead?
> 
> Define something like:
> typedef u64 __attribute__((aligned(4))) u64_align32;
> and then use it for the 64bit structure members.
> 
Hi, David,

Please kindly help check if the following patch is the right fix or
not. I've verified it to work on my test box.

The reason I use this typedef for all the u64 items because there is no
guarantee that the suspend_performance record is in the end of the
memory, thus it may pollute the others.