[PATCH v7 5/9] iommu/arm-smmu-v3: Make arm_smmu_alloc_cd_ptr()

[Mostafa Saleh]

On Mon, Apr 22, 2024 at 11:20:53AM -0300, Jason Gunthorpe wrote:
> On Fri, Apr 19, 2024 at 09:14:21PM +0000, Mostafa Saleh wrote:
> > Hi Jason,
> > 
> > On Tue, Apr 16, 2024 at 04:28:16PM -0300, Jason Gunthorpe wrote:
> > > Only the attach callers can perform an allocation for the CD table entry,
> > > the other callers must not do so, they do not have the correct locking and
> > > they cannot sleep. Split up the functions so this is clear.
> > > 
> > > arm_smmu_get_cd_ptr() will return pointer to a CD table entry without
> > > doing any kind of allocation.
> > > 
> > > arm_smmu_alloc_cd_ptr() will allocate the table and any required
> > > leaf.
> > > 
> > > A following patch will add lockdep assertions to arm_smmu_alloc_cd_ptr()
> > > once the restructuring is completed and arm_smmu_alloc_cd_ptr() is never
> > > called in the wrong context.
> > > 
> > > Signed-off-by: Jason Gunthorpe <jgg at nvidia.com>
> > > ---
> > >  drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.c | 61 +++++++++++++--------
> > >  1 file changed, 39 insertions(+), 22 deletions(-)
> > > 
> > > diff --git a/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.c b/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.c
> > > index f3df1ec8d258dc..a0d1237272936f 100644
> > > --- a/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.c
> > > +++ b/drivers/iommu/arm/arm-smmu-v3/arm-smmu-v3.c
> > > @@ -98,6 +98,7 @@ static struct arm_smmu_option_prop arm_smmu_options[] = {
> > >  
> > >  static int arm_smmu_domain_finalise(struct arm_smmu_domain *smmu_domain,
> > >  				    struct arm_smmu_device *smmu);
> > > +static int arm_smmu_alloc_cd_tables(struct arm_smmu_master *master);
> > >  
> > >  static void parse_driver_options(struct arm_smmu_device *smmu)
> > >  {
> > > @@ -1207,29 +1208,51 @@ static void arm_smmu_write_cd_l1_desc(__le64 *dst,
> > >  struct arm_smmu_cd *arm_smmu_get_cd_ptr(struct arm_smmu_master *master,
> > >  					u32 ssid)
> > >  {
> > > -	__le64 *l1ptr;
> > > -	unsigned int idx;
> > >  	struct arm_smmu_l1_ctx_desc *l1_desc;
> > > -	struct arm_smmu_device *smmu = master->smmu;
> > >  	struct arm_smmu_ctx_desc_cfg *cd_table = &master->cd_table;
> > >  
> > > +	if (!cd_table->cdtab)
> > > +		return NULL;
> > > +
> > >  	if (cd_table->s1fmt == STRTAB_STE_0_S1FMT_LINEAR)
> > >  		return (struct arm_smmu_cd *)(cd_table->cdtab +
> > >  					      ssid * CTXDESC_CD_DWORDS);
> > >  
> > > -	idx = ssid >> CTXDESC_SPLIT;
> > > -	l1_desc = &cd_table->l1_desc[idx];
> > > -	if (!l1_desc->l2ptr) {
> > > -		if (arm_smmu_alloc_cd_leaf_table(smmu, l1_desc))
> > > -			return NULL;
> > > +	l1_desc = &cd_table->l1_desc[ssid / CTXDESC_L2_ENTRIES];
> > 
> > These operations used to be shift and bit masking which made sense as it does
> > what hardware does, is there any reason you changed it to division and modulo?
> > I checked the disassembly and gcc does the right thing as constants are power
> > of 2, but I am just curious.
> 
> I generally prefer the clarity and succinctness of / and % instead of
> hacking up bit operations that the compiler will generate
> automatically anyhow.
> 
> If bit extractions should be used it is better to wrap it in
> FIELD_GET() than open code it..
> 
> > > +static struct arm_smmu_cd *arm_smmu_alloc_cd_ptr(struct arm_smmu_master *master,
> > > +						 u32 ssid)
> > > +{
> > > +	struct arm_smmu_ctx_desc_cfg *cd_table = &master->cd_table;
> > > +	struct arm_smmu_device *smmu = master->smmu;
> > > +
> > > +	if (!cd_table->cdtab) {
> > > +		if (arm_smmu_alloc_cd_tables(master))
> > > +			return NULL;
> > >  	}
> > > -	idx = ssid & (CTXDESC_L2_ENTRIES - 1);
> > > -	return &l1_desc->l2ptr[idx];
> > > +
> > > +	if (cd_table->s1fmt == STRTAB_STE_0_S1FMT_64K_L2) {
> > > +		unsigned int idx = ssid >> CTXDESC_SPLIT;
> > 
> > Ok, now it’s a shift, I think we should be consistent with how we
> > calculate the index.
> 
> Sure. Change that to / will make CTXDESC_SPLIT unused except in
> computing CTXDESC_L2_ENTRIES so that can be simplified too:
> 
> -#define CTXDESC_SPLIT                  10
> -#define CTXDESC_L2_ENTRIES             (1 << CTXDESC_SPLIT)
> +#define CTXDESC_L2_ENTRIES             1024
> 

Sounds good, I don’t think it matters much as long as its consistent, but
anyway the split is defined by the spec to be either 6, 8 or 10.
So split size has to be a power of 2.

> 
> > > @@ -1357,7 +1380,7 @@ int arm_smmu_write_ctx_desc(struct arm_smmu_master *master, int ssid,
> > >  	if (WARN_ON(ssid >= (1 << cd_table->s1cdmax)))
> > >  		return -E2BIG;
> > >  
> > > -	cd_table_entry = arm_smmu_get_cd_ptr(master, ssid);
> > > +	cd_table_entry = arm_smmu_alloc_cd_ptr(master, ssid);
> > 
> > The only path allocates the main table is “arm_smmu_attach_dev”,
> 
> There are two places that allocate the leaf, arm_smmu_attach_dev()
> (for the RID) and arm_smmu_sva_set_dev_pasid() (for a PASID)
> 
> At this moment all the paths are relying on the above to allocate the
> leaf. The next patch makes arm_smmu_attach_dev() allocate the leaf
> itself. A few more patches also makes the PASID path allocate the leaf
> itself, when the above is removed.
> 
> > I guess it would be more robust to leave that as is and have 2
> > versions of get_cd, one that allocates leaf and one that is not
> > allocating, what do you think?
> 
> I'm not sure what you are asking? We have two versions. One is called
> alloc and one is called get. That have different locking requirements
> on the caller so they have different names. I would not call them both
> get?
> 

My point is that arm_smmu_alloc_cd_ptr() doesn’t only allocate the leaf,
but also the L1 through arm_smmu_alloc_cd_tables()

IMO, arm_smmu_alloc_cd_ptr() should only allocate leafs. And inside
arm_smmu_attach_dev() it calls arm_smmu_alloc_cd_tables().
This makes it clear which path is expected to allocate the L1 table.

And arm_smmu_get_cd_ptr() will remain as is.

Thanks,
Mostafa

> Thanks,
> Jason