| Message ID | 20191022125502.12495-2-laurentiu.tudor@nxp.com |
|---|---|
| State | Not Applicable |
| Delegated to: | David Miller |
| Headers | show |
| Series | dma-mapping: introduce a new dma api | expand |
On 22/10/2019 13:55, Laurentiu Tudor wrote: > From: Laurentiu Tudor <laurentiu.tudor@nxp.com> > > Introduce a new dma map op called dma_addr_to_phys_addr() that converts > a dma address to the physical address backing it up and add wrapper for > it. I'd really love it if there was a name which could encapsulate that this is *only* for extreme special cases of constrained descriptors/pagetable entries/etc. where there's simply no practical way to keep track of a CPU address alongside the DMA address, and the only option is this potentially-arbitrarily-complex operation (I mean, on some systems it may end up taking locks and poking hardware). Either way it's tricky - much as I don't like adding an interface which is ripe for drivers to misuse, I also really don't want hacks like bdf95923086f shoved into other APIs to compensate, so on balance I'd probably consider this proposal ever so slightly the lesser evil. > Signed-off-by: Laurentiu Tudor <laurentiu.tudor@nxp.com> > --- > include/linux/dma-mapping.h | 21 +++++++++++++++++++++ > 1 file changed, 21 insertions(+) > > diff --git a/include/linux/dma-mapping.h b/include/linux/dma-mapping.h > index 4a1c4fca475a..5965d159c9a9 100644 > --- a/include/linux/dma-mapping.h > +++ b/include/linux/dma-mapping.h > @@ -132,6 +132,8 @@ struct dma_map_ops { > u64 (*get_required_mask)(struct device *dev); > size_t (*max_mapping_size)(struct device *dev); > unsigned long (*get_merge_boundary)(struct device *dev); > + phys_addr_t (*dma_addr_to_phys_addr)(struct device *dev, > + dma_addr_t dma_handle); I'd be inclined to name the internal callback something a bit snappier like .get_phys_addr. > }; > > #define DMA_MAPPING_ERROR (~(dma_addr_t)0) > @@ -442,6 +444,19 @@ static inline int dma_mapping_error(struct device *dev, dma_addr_t dma_addr) > return 0; > } > > +static inline phys_addr_t dma_addr_to_phys_addr(struct device *dev, > + dma_addr_t dma_handle) > +{ > + const struct dma_map_ops *ops = get_dma_ops(dev); > + > + if (dma_is_direct(ops)) > + return (phys_addr_t)dma_handle; Well that's not right, is it - remember why you had that namespace collision? ;) Robin. > + else if (ops->dma_addr_to_phys_addr) > + return ops->dma_addr_to_phys_addr(dev, dma_handle); > + > + return 0; > +} > + > void *dma_alloc_attrs(struct device *dev, size_t size, dma_addr_t *dma_handle, > gfp_t flag, unsigned long attrs); > void dma_free_attrs(struct device *dev, size_t size, void *cpu_addr, > @@ -578,6 +593,12 @@ static inline unsigned long dma_get_merge_boundary(struct device *dev) > { > return 0; > } > + > +static inline phys_addr_t dma_addr_to_phys_addr(struct device *dev, > + dma_addr_t dma_handle) > +{ > + return 0; > +} > #endif /* CONFIG_HAS_DMA */ > > static inline dma_addr_t dma_map_single_attrs(struct device *dev, void *ptr, >
On 22.10.2019 16:25, Robin Murphy wrote: > On 22/10/2019 13:55, Laurentiu Tudor wrote: >> From: Laurentiu Tudor <laurentiu.tudor@nxp.com> >> >> Introduce a new dma map op called dma_addr_to_phys_addr() that converts >> a dma address to the physical address backing it up and add wrapper for >> it. > > I'd really love it if there was a name which could encapsulate that this > is *only* for extreme special cases of constrained descriptors/pagetable > entries/etc. where there's simply no practical way to keep track of a > CPU address alongside the DMA address, and the only option is this > potentially-arbitrarily-complex operation (I mean, on some systems it > may end up taking locks and poking hardware). > > Either way it's tricky - much as I don't like adding an interface which > is ripe for drivers to misuse, I also really don't want hacks like > bdf95923086f shoved into other APIs to compensate, so on balance I'd > probably consider this proposal ever so slightly the lesser evil. > >> Signed-off-by: Laurentiu Tudor <laurentiu.tudor@nxp.com> >> --- >> include/linux/dma-mapping.h | 21 +++++++++++++++++++++ >> 1 file changed, 21 insertions(+) >> >> diff --git a/include/linux/dma-mapping.h b/include/linux/dma-mapping.h >> index 4a1c4fca475a..5965d159c9a9 100644 >> --- a/include/linux/dma-mapping.h >> +++ b/include/linux/dma-mapping.h >> @@ -132,6 +132,8 @@ struct dma_map_ops { >> u64 (*get_required_mask)(struct device *dev); >> size_t (*max_mapping_size)(struct device *dev); >> unsigned long (*get_merge_boundary)(struct device *dev); >> + phys_addr_t (*dma_addr_to_phys_addr)(struct device *dev, >> + dma_addr_t dma_handle); > > I'd be inclined to name the internal callback something a bit snappier > like .get_phys_addr. Alright. Want me to also rename the wrapper to something like dma_get_phys_addr()? Sounds a bit nicer to me. >> }; >> #define DMA_MAPPING_ERROR (~(dma_addr_t)0) >> @@ -442,6 +444,19 @@ static inline int dma_mapping_error(struct device >> *dev, dma_addr_t dma_addr) >> return 0; >> } >> +static inline phys_addr_t dma_addr_to_phys_addr(struct device *dev, >> + dma_addr_t dma_handle) >> +{ >> + const struct dma_map_ops *ops = get_dma_ops(dev); >> + >> + if (dma_is_direct(ops)) >> + return (phys_addr_t)dma_handle; > > Well that's not right, is it - remember why you had that namespace > collision? ;) > Ugh, correct. Don't know what I was thinking. Will rework the check. --- Thanks & Best Regards, Laurentiu
Hi Robin, On 22.10.2019 16:25, Robin Murphy wrote: > On 22/10/2019 13:55, Laurentiu Tudor wrote: >> From: Laurentiu Tudor <laurentiu.tudor@nxp.com> >> >> Introduce a new dma map op called dma_addr_to_phys_addr() that converts >> a dma address to the physical address backing it up and add wrapper for >> it. > > I'd really love it if there was a name which could encapsulate that this > is *only* for extreme special cases of constrained descriptors/pagetable > entries/etc. where there's simply no practical way to keep track of a > CPU address alongside the DMA address, and the only option is this > potentially-arbitrarily-complex operation (I mean, on some systems it > may end up taking locks and poking hardware). > > Either way it's tricky - much as I don't like adding an interface which > is ripe for drivers to misuse, I also really don't want hacks like > bdf95923086f shoved into other APIs to compensate, so on balance I'd > probably consider this proposal ever so slightly the lesser evil. We had an internal discussion over these points you are raising and Madalin (cc-ed) came up with another idea: instead of adding this prone to misuse api how about experimenting with a new dma unmap and dma sync variants that would return the physical address by calling the newly introduced dma map op. Something along these lines: * phys_addr_t dma_unmap_page_ret_phys(...) * phys_addr_t dma_unmap_single_ret_phys(...) * phys_addr_t dma_sync_single_for_cpu_ret_phys(...) I'm thinking that this proposal should reduce the risks opened by the initial variant. Please let me know what you think. --- Thanks & Best Regards, Laurentiu
On Wed, Oct 23, 2019 at 11:53:41AM +0000, Laurentiu Tudor wrote: > We had an internal discussion over these points you are raising and > Madalin (cc-ed) came up with another idea: instead of adding this prone > to misuse api how about experimenting with a new dma unmap and dma sync > variants that would return the physical address by calling the newly > introduced dma map op. Something along these lines: > * phys_addr_t dma_unmap_page_ret_phys(...) > * phys_addr_t dma_unmap_single_ret_phys(...) > * phys_addr_t dma_sync_single_for_cpu_ret_phys(...) > I'm thinking that this proposal should reduce the risks opened by the > initial variant. > Please let me know what you think. I'm not sure what the ret is supposed to mean, but I generally like that idea better. We also need to make sure there is an easy way to figure out if these APIs are available, as they generally aren't for any non-IOMMU API IOMMU drivers.
On 24.10.2019 05:01, hch@lst.de wrote: > On Wed, Oct 23, 2019 at 11:53:41AM +0000, Laurentiu Tudor wrote: >> We had an internal discussion over these points you are raising and >> Madalin (cc-ed) came up with another idea: instead of adding this prone >> to misuse api how about experimenting with a new dma unmap and dma sync >> variants that would return the physical address by calling the newly >> introduced dma map op. Something along these lines: >> * phys_addr_t dma_unmap_page_ret_phys(...) >> * phys_addr_t dma_unmap_single_ret_phys(...) >> * phys_addr_t dma_sync_single_for_cpu_ret_phys(...) >> I'm thinking that this proposal should reduce the risks opened by the >> initial variant. >> Please let me know what you think. > > I'm not sure what the ret is supposed to mean, but I generally like > that idea better. It was supposed to be short for "return" but given that I'm not good at naming stuff I'll just drop it. > We also need to make sure there is an easy way > to figure out if these APIs are available, as they generally aren't > for any non-IOMMU API IOMMU drivers. I was really hoping to manage making them as generic as possible but anyway, I'll start working on a PoC and see how it turns out. This will probably happen sometime next next week as the following week I'll be traveling to a conference. --- Best Regards, Laurentiu
On 2019-10-24 8:49 am, Laurentiu Tudor wrote: > > > On 24.10.2019 05:01, hch@lst.de wrote: >> On Wed, Oct 23, 2019 at 11:53:41AM +0000, Laurentiu Tudor wrote: >>> We had an internal discussion over these points you are raising and >>> Madalin (cc-ed) came up with another idea: instead of adding this prone >>> to misuse api how about experimenting with a new dma unmap and dma sync >>> variants that would return the physical address by calling the newly >>> introduced dma map op. Something along these lines: >>> * phys_addr_t dma_unmap_page_ret_phys(...) >>> * phys_addr_t dma_unmap_single_ret_phys(...) >>> * phys_addr_t dma_sync_single_for_cpu_ret_phys(...) >>> I'm thinking that this proposal should reduce the risks opened by the >>> initial variant. >>> Please let me know what you think. >> >> I'm not sure what the ret is supposed to mean, but I generally like >> that idea better. > > It was supposed to be short for "return" but given that I'm not good at > naming stuff I'll just drop it. Hmm, how about something like "dma_unmap_*_desc" for the context of the mapped DMA address also being used as a descriptor token? >> We also need to make sure there is an easy way >> to figure out if these APIs are available, as they generally aren't >> for any non-IOMMU API IOMMU drivers. > > I was really hoping to manage making them as generic as possible but > anyway, I'll start working on a PoC and see how it turns out. This will > probably happen sometime next next week as the following week I'll be > traveling to a conference. AFAICS, even a full implementation of these APIs would have to be capable of returning an indication that there is no valid physical address - e.g. if unmap is called with a bogus DMA address that was never mapped. At that point there'sseemingly no problem just implementing the trivial case on top of any existing unmap/sync callbacks for everyone. I'd imagine that drivers which want this aren't likely to run on the older architectures where the weird IOMMUs live, so they could probably just always treat failure as unexpected and fatal either way. In fact, I'm now wondering whether it's likely to be common that users want the physical address specifically, or whether it would make sense to return the original VA/page, both for symmetry with the corresponding map calls and for the ease of being able to return NULL when necessary. Robin.
On 24.10.2019 14:04, Robin Murphy wrote: > On 2019-10-24 8:49 am, Laurentiu Tudor wrote: >> >> >> On 24.10.2019 05:01, hch@lst.de wrote: >>> On Wed, Oct 23, 2019 at 11:53:41AM +0000, Laurentiu Tudor wrote: >>>> We had an internal discussion over these points you are raising and >>>> Madalin (cc-ed) came up with another idea: instead of adding this prone >>>> to misuse api how about experimenting with a new dma unmap and dma sync >>>> variants that would return the physical address by calling the newly >>>> introduced dma map op. Something along these lines: >>>> * phys_addr_t dma_unmap_page_ret_phys(...) >>>> * phys_addr_t dma_unmap_single_ret_phys(...) >>>> * phys_addr_t dma_sync_single_for_cpu_ret_phys(...) >>>> I'm thinking that this proposal should reduce the risks opened by the >>>> initial variant. >>>> Please let me know what you think. >>> >>> I'm not sure what the ret is supposed to mean, but I generally like >>> that idea better. >> >> It was supposed to be short for "return" but given that I'm not good at >> naming stuff I'll just drop it. > > Hmm, how about something like "dma_unmap_*_desc" for the context of the > mapped DMA address also being used as a descriptor token? Alright. >>> We also need to make sure there is an easy way >>> to figure out if these APIs are available, as they generally aren't >>> for any non-IOMMU API IOMMU drivers. >> >> I was really hoping to manage making them as generic as possible but >> anyway, I'll start working on a PoC and see how it turns out. This will >> probably happen sometime next next week as the following week I'll be >> traveling to a conference. > > AFAICS, even a full implementation of these APIs would have to be > capable of returning an indication that there is no valid physical > address - e.g. if unmap is called with a bogus DMA address that was > never mapped. At that point there'sseemingly no problem just > implementing the trivial case on top of any existing unmap/sync > callbacks for everyone. I'd imagine that drivers which want this aren't > likely to run on the older architectures where the weird IOMMUs live, so > they could probably just always treat failure as unexpected and fatal > either way. > > In fact, I'm now wondering whether it's likely to be common that users > want the physical address specifically, or whether it would make sense > to return the original VA/page, both for symmetry with the corresponding > map calls and for the ease of being able to return NULL when necessary. That's sounds wonderful as it should make the code leaner in the drivers. --- Best Regards, Laurentiu
diff --git a/include/linux/dma-mapping.h b/include/linux/dma-mapping.h index 4a1c4fca475a..5965d159c9a9 100644 --- a/include/linux/dma-mapping.h +++ b/include/linux/dma-mapping.h @@ -132,6 +132,8 @@ struct dma_map_ops { u64 (*get_required_mask)(struct device *dev); size_t (*max_mapping_size)(struct device *dev); unsigned long (*get_merge_boundary)(struct device *dev); + phys_addr_t (*dma_addr_to_phys_addr)(struct device *dev, + dma_addr_t dma_handle); }; #define DMA_MAPPING_ERROR (~(dma_addr_t)0) @@ -442,6 +444,19 @@ static inline int dma_mapping_error(struct device *dev, dma_addr_t dma_addr) return 0; } +static inline phys_addr_t dma_addr_to_phys_addr(struct device *dev, + dma_addr_t dma_handle) +{ + const struct dma_map_ops *ops = get_dma_ops(dev); + + if (dma_is_direct(ops)) + return (phys_addr_t)dma_handle; + else if (ops->dma_addr_to_phys_addr) + return ops->dma_addr_to_phys_addr(dev, dma_handle); + + return 0; +} + void *dma_alloc_attrs(struct device *dev, size_t size, dma_addr_t *dma_handle, gfp_t flag, unsigned long attrs); void dma_free_attrs(struct device *dev, size_t size, void *cpu_addr, @@ -578,6 +593,12 @@ static inline unsigned long dma_get_merge_boundary(struct device *dev) { return 0; } + +static inline phys_addr_t dma_addr_to_phys_addr(struct device *dev, + dma_addr_t dma_handle) +{ + return 0; +} #endif /* CONFIG_HAS_DMA */ static inline dma_addr_t dma_map_single_attrs(struct device *dev, void *ptr,