Re: [PATCH v9 19/27] dmaengine: dw-edma: Use non-atomic io-64 methods

[Serge Semin]

On Fri, Jan 20, 2023 at 06:54:01PM -0600, Bjorn Helgaas wrote:
> On Fri, Jan 13, 2023 at 08:14:01PM +0300, Serge Semin wrote:
> > Instead of splitting the 64-bits IOs up into two 32-bits ones it's
> > possible to use the already available non-atomic readq/writeq methods
> > implemented exactly for such cases. They are defined in the dedicated
> > header files io-64-nonatomic-lo-hi.h/io-64-nonatomic-hi-lo.h. So in case
> > if the 64-bits readq/writeq methods are unavailable on some platforms at
> > consideration, the corresponding drivers can have any of these headers
> > included and stop locally re-implementing the 64-bits IO accessors taking
> > into account the non-atomic nature of the included methods. Let's do that
> > in the DW eDMA driver too. Note by doing so we can discard the
> > CONFIG_64BIT config ifdefs from the code.
> > 
> > Signed-off-by: Serge Semin <Sergey.Semin@xxxxxxxxxxxxxxxxxxxx>
> > Reviewed-by: Manivannan Sadhasivam <manivannan.sadhasivam@xxxxxxxxxx>
> > Tested-by: Manivannan Sadhasivam <manivannan.sadhasivam@xxxxxxxxxx>
> > Acked-by: Vinod Koul <vkoul@xxxxxxxxxx>
> > ---
> >  drivers/dma/dw-edma/dw-edma-v0-core.c | 55 +++++++++------------------
> >  1 file changed, 18 insertions(+), 37 deletions(-)
> > 
> > diff --git a/drivers/dma/dw-edma/dw-edma-v0-core.c b/drivers/dma/dw-edma/dw-edma-v0-core.c
> > index 66f296daac5a..51a34b43434c 100644
> > --- a/drivers/dma/dw-edma/dw-edma-v0-core.c
> > +++ b/drivers/dma/dw-edma/dw-edma-v0-core.c
> > @@ -8,6 +8,8 @@
> >  
> >  #include <linux/bitfield.h>
> >  
> > +#include <linux/io-64-nonatomic-lo-hi.h>
> > +
> >  #include "dw-edma-core.h"
> >  #include "dw-edma-v0-core.h"
> >  #include "dw-edma-v0-regs.h"
> > @@ -53,8 +55,6 @@ static inline struct dw_edma_v0_regs __iomem *__dw_regs(struct dw_edma *dw)
> >  		SET_32(dw, rd_##name, value);		\
> >  	} while (0)
> >  
> > -#ifdef CONFIG_64BIT
> > -
> >  #define SET_64(dw, name, value)				\
> >  	writeq(value, &(__dw_regs(dw)->name))
> >  
> > @@ -80,8 +80,6 @@ static inline struct dw_edma_v0_regs __iomem *__dw_regs(struct dw_edma *dw)
> >  		SET_64(dw, rd_##name, value);		\
> >  	} while (0)
> >  
> > -#endif /* CONFIG_64BIT */
> 

> Great to get rid of these #ifdefs!
> 
> Am I missing something?  It looks like SET_64 is used only by
> SET_RW_64 and SET_BOTH_64, and neither of *them is used at all.
> 
> Similarly for GET_64 and GET_RW_64.
> 
> So maybe we could get rid of everything inside the #ifdefs as well?

Even though these macros are indeed unused in the driver they are
still a part of the DW eDMA CSRs access interface. In particular they
are supposed to be used to access the 64-bit registers declared in the
dw_edma_v0_regs, dw_edma_v0_unroll and dw_edma_v0_ch_regs structures.
So until the interface is converted to a more preferable direct MMIO
usage without any packed-structures I'd rather leave these macros
be.

> 
> >  #define SET_COMPAT(dw, name, value)			\
> >  	writel(value, &(__dw_regs(dw)->type.unroll.name))
> >  
> > @@ -164,14 +162,13 @@ static inline u32 readl_ch(struct dw_edma *dw, enum dw_edma_dir dir, u16 ch,
> >  #define SET_LL_32(ll, value) \
> >  	writel(value, ll)
> >  
> > -#ifdef CONFIG_64BIT
> > -
> >  static inline void writeq_ch(struct dw_edma *dw, enum dw_edma_dir dir, u16 ch,
> >  			     u64 value, void __iomem *addr)
> >  {
> > +	unsigned long flags;
> > +
> >  	if (dw->chip->mf == EDMA_MF_EDMA_LEGACY) {
> >  		u32 viewport_sel;
> > -		unsigned long flags;
> >  
> >  		raw_spin_lock_irqsave(&dw->lock, flags);
> >  
> > @@ -181,22 +178,22 @@ static inline void writeq_ch(struct dw_edma *dw, enum dw_edma_dir dir, u16 ch,
> >  
> >  		writel(viewport_sel,
> >  		       &(__dw_regs(dw)->type.legacy.viewport_sel));
> > -		writeq(value, addr);
> > +	}
> > +
> > +	writeq(value, addr);
> >  
> > +	if (dw->chip->mf == EDMA_MF_EDMA_LEGACY)
> >  		raw_spin_unlock_irqrestore(&dw->lock, flags);
> > -	} else {
> > -		writeq(value, addr);
> > -	}
> 

> This is basically a cosmetic change unrelated to the commit log.  I
> don't really object to the change, although I think it's of dubious
> value to remove the repetition of the writeq() at the cost of adding
> another "if" and unlock.
> 

The denoted change is basically a leftover of the originally more
complex modification:
https://lore.kernel.org/linux-pci/20220503225104.12108-21-Sergey.Semin@xxxxxxxxxxxxxxxxxxxx
for which the update made more sense. But then the corresponding flag
was dropped from the Frank' patchset and I had to reduce the patch
to the code above.

Though I agree with you in both aspects. Indeed the value of the
change is questionable and it isn't related to the commit log.

> Lorenzo already applied this, so it's OK as-is unless you think it's
> worth reworking to drop the unused stuff mentioned above, in which
> case this rearrangement could be moved to a separate patch to make
> both of them more focused.

As I said above I'd rather leave the 64-bit accessors be until the
packed structure-based interface is removed from the driver.

Regarding the QWORD accessors update. If @Lorenzo agrees to replace
the already applied v9 patchset with a new one I can resubmit the
series with no rearrangement above.

-Serge(y)

> 
> Bjorn