| Message ID | 20210225143339.3693838-1-arnd@kernel.org (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | drm/amd/display: fix 64-bit integer division | expand |
On Thu, Feb 25, 2021 at 3:33 PM Arnd Bergmann <arnd@kernel.org> wrote: > > From: Arnd Bergmann <arnd@arndb.de> > > The new display synchronization code caused a regression > on all 32-bit architectures: > > ld.lld: error: undefined symbol: __aeabi_uldivmod > >>> referenced by dce_clock_source.c > >>> gpu/drm/amd/display/dc/dce/dce_clock_source.o:(get_pixel_clk_frequency_100hz) in archive drivers/built-in.a > > ld.lld: error: undefined symbol: __aeabi_ldivmod > >>> referenced by dc_resource.c > >>> gpu/drm/amd/display/dc/core/dc_resource.o:(resource_are_vblanks_synchronizable) in archive drivers/built-in.a > >>> referenced by dc_resource.c > >>> gpu/drm/amd/display/dc/core/dc_resource.o:(resource_are_vblanks_synchronizable) in archive drivers/built-in.a > >>> referenced by dc_resource.c > >>> gpu/drm/amd/display/dc/core/dc_resource.o:(resource_are_vblanks_synchronizable) in archive drivers/built-in.a > > This is not a fast path, so the use of an explicit div_u64/div_s64 > seems appropriate. I found two more instances: >>> referenced by dcn20_optc.c >>> gpu/drm/amd/display/dc/dcn20/dcn20_optc.o:(optc2_align_vblanks) in archive drivers/built-in.a >>> referenced by dcn10_hw_sequencer.c >>> gpu/drm/amd/display/dc/dcn10/dcn10_hw_sequencer.o:(reduceSizeAndFraction) in archive drivers/built-in.a I have patches for both, but will let the randconfig build box keep working on it over night to see if there are any others. Let me know if you want a combined patch or one per file once there are no more regressions. Arnd
[AMD Official Use Only - Internal Distribution Only]
Hi Arnd,
I have all the patches ready and I have tested them with the feature/platform I'm working on and Bindu helped to test the 32bit build.
I'm in process of submitting the latest change.
Thanks,
Vladimir.
On 2021-02-25, 4:36 PM, "Arnd Bergmann" <arnd@kernel.org> wrote:
On Thu, Feb 25, 2021 at 3:33 PM Arnd Bergmann <arnd@kernel.org> wrote:
>
> From: Arnd Bergmann <arnd@arndb.de>
>
> The new display synchronization code caused a regression
> on all 32-bit architectures:
>
> ld.lld: error: undefined symbol: __aeabi_uldivmod
> >>> referenced by dce_clock_source.c
> >>> gpu/drm/amd/display/dc/dce/dce_clock_source.o:(get_pixel_clk_frequency_100hz) in archive drivers/built-in.a
>
> ld.lld: error: undefined symbol: __aeabi_ldivmod
> >>> referenced by dc_resource.c
> >>> gpu/drm/amd/display/dc/core/dc_resource.o:(resource_are_vblanks_synchronizable) in archive drivers/built-in.a
> >>> referenced by dc_resource.c
> >>> gpu/drm/amd/display/dc/core/dc_resource.o:(resource_are_vblanks_synchronizable) in archive drivers/built-in.a
> >>> referenced by dc_resource.c
> >>> gpu/drm/amd/display/dc/core/dc_resource.o:(resource_are_vblanks_synchronizable) in archive drivers/built-in.a
>
> This is not a fast path, so the use of an explicit div_u64/div_s64
> seems appropriate.
I found two more instances:
>>> referenced by dcn20_optc.c
>>> gpu/drm/amd/display/dc/dcn20/dcn20_optc.o:(optc2_align_vblanks) in archive drivers/built-in.a
>>> referenced by dcn10_hw_sequencer.c
>>> gpu/drm/amd/display/dc/dcn10/dcn10_hw_sequencer.o:(reduceSizeAndFraction) in archive drivers/built-in.a
I have patches for both, but will let the randconfig build box keep working
on it over night to see if there are any others. Let me know if you want a
combined patch or one per file once there are no more regressions.
Arnd
diff --git a/drivers/gpu/drm/amd/display/dc/core/dc_resource.c b/drivers/gpu/drm/amd/display/dc/core/dc_resource.c index 0241c9d96d7a..49214c59c836 100644 --- a/drivers/gpu/drm/amd/display/dc/core/dc_resource.c +++ b/drivers/gpu/drm/amd/display/dc/core/dc_resource.c @@ -441,15 +441,15 @@ bool resource_are_vblanks_synchronizable( if (stream2->timing.pix_clk_100hz*100/stream2->timing.h_total/ stream2->timing.v_total > 60) return false; - frame_time_diff = (int64_t)10000 * + frame_time_diff = div_s64(10000ll * stream1->timing.h_total * stream1->timing.v_total * - stream2->timing.pix_clk_100hz / - stream1->timing.pix_clk_100hz / - stream2->timing.h_total / - stream2->timing.v_total; + stream2->timing.pix_clk_100hz, + stream1->timing.pix_clk_100hz * + stream2->timing.h_total * + stream2->timing.v_total); for (i = 0; i < rr_count; i++) { - int64_t diff = (frame_time_diff * base60_refresh_rates[i]) / 10 - 10000; + int64_t diff = div_s64(frame_time_diff * base60_refresh_rates[i], 10) - 10000; if (diff < 0) diff = -diff; diff --git a/drivers/gpu/drm/amd/display/dc/dce/dce_clock_source.c b/drivers/gpu/drm/amd/display/dc/dce/dce_clock_source.c index 6f47f9bab5ee..85ed6f2c9647 100644 --- a/drivers/gpu/drm/amd/display/dc/dce/dce_clock_source.c +++ b/drivers/gpu/drm/amd/display/dc/dce/dce_clock_source.c @@ -1013,9 +1013,9 @@ static bool get_pixel_clk_frequency_100hz( * not be programmed equal to DPREFCLK */ modulo_hz = REG_READ(MODULO[inst]); - *pixel_clk_khz = ((uint64_t)clock_hz* - clock_source->ctx->dc->clk_mgr->dprefclk_khz*10)/ - modulo_hz; + *pixel_clk_khz = div_u64((uint64_t)clock_hz * 10 * + clock_source->ctx->dc->clk_mgr->dprefclk_khz, + modulo_hz); } else { /* NOTE: There is agreement with VBIOS here that MODULO is * programmed equal to DPREFCLK, in which case PHASE will be