On Thu, Jan 30, 2020 at 8:44 AM Patrice CHOTARD <patrice.chotard@xxxxxx> wrote:
>
> Hi Rob, Tom, David
>
> On 1/28/20 3:08 PM, Rob Herring wrote:
> > On Tue, Jan 28, 2020 at 7:43 AM Tom Rini <trini@xxxxxxxxxxxx> wrote:
> >> On Tue, Jan 28, 2020 at 07:59:18PM +1100, David Gibson wrote:
> >>> On Mon, Jan 27, 2020 at 10:04:34AM -0500, Tom Rini wrote:
> >>>> On Mon, Jan 27, 2020 at 02:23:51PM +1100, David Gibson wrote:
> >>>>> On Thu, Jan 23, 2020 at 07:16:50AM -0500, Tom Rini wrote:
> >>>>>> On Thu, Jan 23, 2020 at 08:14:40PM +1100, David Gibson wrote:
> >>>>>>> On Fri, Jan 17, 2020 at 10:31:06AM -0500, Tom Rini wrote:
> >>>>>>>> 6dcb8ba4 "libfdt: Add helpers for accessing unaligned words" introduced
> >>>>>>>> changes to support unaligned reads for ARM platforms and 11738cf01f15
> >>>>>>>> "libfdt: Don't use memcpy to handle unaligned reads on ARM" improved the
> >>>>>>>> performance of these helpers.
> >>>>>>>>
> >>>>>>>> Ultimately however, these helpers should not exist. Unaligned access
> >>>>>>>> only occurs when images are forced out of alignment by the user. This
> >>>>>>>> unalignment is not supported and introduces problems later on as other
> >>>>>>>> parts of the system image are unaligned and they too require alignment.
> >>>>>>>>
> >>>>>>>> Revert both of these changes.
> >>>>>>>>
> >>>>>>>> Signed-off-by: Tom Rini <trini@xxxxxxxxxxxx>
> >>>>>>>> ---
> >>>>>>>> By way of a little more explanation, looking at the archives it seems
> >>>>>>>> that the initial bug reporter said that they had a platform that was
> >>>>>>>> using U-Boot and had the "fdt_high=0xffffffff" set in the environment.
> >>>>>>>> What that does is to tell U-Boot to not do any of the sanity checks and
> >>>>>>>> relocation to ensure alignment that it would normally do because the
> >>>>>>>> user knows best. This later came up on the U-Boot list as once the DTB
> >>>>>>>> was loaded, Linux is unhappy because it demands correct alignment.
> >>>>>>>>
> >>>>>>>> I only realized libfdt had introduced changes here when it was reported
> >>>>>>>> that boot time had gotten much slower once we merged this change in. It
> >>>>>>>> would be best to just drop it.
> >>>>>>> Hmm. I'm not sure about this. The commit message makes a case for
> >>>>>>> why the unaligned handling isn't necessary, but not a case for why
> >>>>>>> it's bad. Even if handling an unaligned tree isn't a common case,
> >>>>>>> isn't it better to be able to than not?
> >>>>>>>
> >>>>>>> I gather from the previous discussion that there's a significant
> >>>>>>> performance impact, but that rationale needs to go into the commit
> >>>>>>> message for posterity.
> >>>>>> I wanted to emphasize that the code simply isn't ever needed, not that
> >>>>>> it's a performance problem. A performance problem implies that we would
> >>>>>> keep this, if it was fast enough. That's why people noticed it (it
> >>>>>> slows things down to an unusable level). But it's functionally wrong.
> >>>>>>
> >>>>>>> [snip]
> >>>>>>>> diff --git a/libfdt/libfdt.h b/libfdt/libfdt.h
> >>>>>>>> index fc4c4962a01c..d4ebe915cf46 100644
> >>>>>>>> --- a/libfdt/libfdt.h
> >>>>>>>> +++ b/libfdt/libfdt.h
> >>>>>>>> @@ -117,23 +117,6 @@ static inline void *fdt_offset_ptr_w(void *fdt, int offset, int checklen)
> >>>>>>>>
> >>>>>>>> uint32_t fdt_next_tag(const void *fdt, int offset, int *nextoffset);
> >>>>>>>>
> >>>>>>>> -/*
> >>>>>>>> - * Alignment helpers:
> >>>>>>>> - * These helpers access words from a device tree blob. They're
> >>>>>>>> - * built to work even with unaligned pointers on platforms (ike
> >>>>>>>> - * ARM) that don't like unaligned loads and stores
> >>>>>>>> - */
> >>>>>>>> -
> >>>>>>>> -static inline uint32_t fdt32_ld(const fdt32_t *p)
> >>>>>>>> -{
> >>>>>>>> - const uint8_t *bp = (const uint8_t *)p;
> >>>>>>>> -
> >>>>>>>> - return ((uint32_t)bp[0] << 24)
> >>>>>>>> - | ((uint32_t)bp[1] << 16)
> >>>>>>>> - | ((uint32_t)bp[2] << 8)
> >>>>>>>> - | bp[3];
> >>>>>>>> -}
> >>>>>>> In particular, I definitely think removing the helpers entirely is a
> >>>>>>> no go. They're now part of the published interface of the library.
> >>>>>> Perhaps "mistakes were made" ? I don't think we need to worry about
> >>>>>> removing an interface here as projects are avoiding upgrading libfdt
> >>>>>> now (TF-A, formerly ATF) or reverting the change (U-Boot).
> >>>>>>
> >>>>>>> Even if they're not used for reading the internal tags, they can be
> >>>>>>> used to load integers from within particular properties. Those are
> >>>>>>> frequently unaligned, since properties generally have packed
> >>>>>>> representations.
> >>>>>> I don't see the relevance. Go back to the initial problem report. It's
> >>>>>> not "I have a new unique platform I'm using libfdt on and I have
> >>>>>> problems". It's "I keep jabbing myself with a rusty nail and now I have
> >>>>>> problems".
> >>>>> The initial report isn't the only relevant thing here. Although it
> >>>>> prompted the change, it wasn't the only consideration in making it.
> >>>>>
> >>>>> There's also two separate questions here:
> >>>>> 1) Do we want byteswapping integer load helpers?
> >>>>> 2) Should those handle unaligned accesses?
> >>>>>
> >>>>> In working out how to address the (as it turns out, non existent)
> >>>>> problem, I realized an abstraction for loading big-endian integers
> >>>>> from the blob would be a useful thing in its own right. I also
> >>>>> realized that it's a useful thing not just inside the libfdt code, but
> >>>>> as an external interface. Callers have always needed to interpret the
> >>>>> contents of individual dt properties, and loading integers from them
> >>>>> is often a part of that.
> >>>>>
> >>>>> I know of people using those fdtXX_ld() helpers right now, so I'm not
> >>>>> going to take them away.
> >>>>>
> >>>>> For the case of external users we absolutely do need to handle
> >>>>> unaligned accesses. There are a number of existing bindings that mix
> >>>>> strings and integers in packed format, in a single encoded property.
> >>>>> So regardless of the alignment of the whole property, we can easily
> >>>>> get unaligned integers in there, and I don't want to expose multiple
> >>>>> different helpers for different cases.
> >>>>>
> >>>>> Now, we don't *have* to use the same helpers for internal use. We
> >>>>> could open code the internal loads, or use a special aligned-only
> >>>>> version inside. But using the existing external helpers is the
> >>>>> obvious and simple choice.
> >>>>>
> >>>>> So, we're back to: I need a case for changing this now, not just a
> >>>>> case for claiming it wasn't needed in the first place.
> >>>> For U-Boot, I'm just going to revert this part of the changes. I would
> >>> That seems reasonable for the interim.
> >>>
> >>>> suggest that you look in to some way to fix the "fast path" here to go
> >>>> back to the previous way of working so that other project can continue
> >>>> to use libfdt as well and when callers need to access these helpers and
> >>>> are not otherwise in the fast path can do so.
> >>>>
> >>>> You're adding visible boot time delay with things the way they exist
> >>>> today. That's not OK.
> >>> That's a fair point, but you need to make that case in the commit
> >>> message and merge request, not just expect me to find and collate the
> >>> arguments from other threads.
> >> If you want me to leave the helpers alone but otherwise revert things, I
> >> can do a v2 and expand the commit message. And perhaps I'm too nice
> >> sometimes then but I do pickup and tweak things that are close enough to
> >> what I want and reword as needed for clarity.
> > Why not just fix the helpers for the aligned case and be done with it:
> >
> > static inline uint32_t fdt32_ld(const fdt32_t *p)
> > {
> > const uint8_t *bp = (const uint8_t *)p;
> >
> > + if (!((unsigned long)p & 0x3))
> > + return fdt32_to_cpu(*p);
> > +
> > return ((uint32_t)bp[0] << 24)
> > | ((uint32_t)bp[1] << 16)
> > | ((uint32_t)bp[2] << 8)
> > | bp[3];
> > }
>
>
> Here are the results, before and after your proposed patch:
>
> tested tag: V2020.04-RC1
>
> STM32MP> bootstage report
> Timer summary in microseconds (12 records):
> Mark Elapsed Stage
> 0 0 reset
> 84,268 84,268 SPL
> 962,921 878,653 end SPL
> 965,800 2,879 board_init_f
> 4,314,348 3,348,548 board_init_r
> 4,863,763 549,415 id=64
> 4,908,759 44,996 id=65
> 4,909,459 700 main_loop
> 5,322,309 412,850 id=175
>
> Accumulated time:
> 83,284 dm_r
> 95,842 dm_spl
> 1,502,020 dm_f
>
>
> -------------------------------------------------------------------------------
>
> tested tag : V2020.04-RC1 + following fdt32_ld patch :
>
> static inline uint32_t fdt32_ld(const fdt32_t *p)
> {
> const uint8_t *bp = (const uint8_t *)p;
>
> + if (!((unsigned long)p & 0x3))
> + return fdt32_to_cpu(*p);
> +
> return ((uint32_t)bp[0] << 24)
> | ((uint32_t)bp[1] << 16)
> | ((uint32_t)bp[2] << 8)
> | bp[3];
> }
>
>
> STM32MP> bootstage report
> Timer summary in microseconds (12 records):
> Mark Elapsed Stage
> 0 0 reset
> 84,264 84,264 SPL
> 959,300 875,036 end SPL
> 962,192 2,892 board_init_f
> 4,310,598 3,348,406 board_init_r
> 4,860,074 549,476 id=64
> 4,905,119 45,045 id=65
> 4,905,819 700 main_loop
> 5,098,636 192,817 id=175
>
> Accumulated time:
> 83,202 dm_r
> 95,252 dm_spl
> 1,501,950 dm_f
>
>
> There is no gain in board_init_r(), the added alignment test is expensive itself.
That's odd. It should be just an AND and a conditional branch added.
Those should be a few cycles compared to tens to hundreds of cycles
for 4 loads instead of 1.
Doesn't u-boot build with -Os typically? Maybe it's not actually inlining.
Rob
