| Message ID | 1490822037-6752-3-git-send-email-roy.pledge@nxp.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
Hi Roy, On 29/03/17 22:13, Roy Pledge wrote: > Use the shared-memory-pool mechanism for frame queue descriptor and > packed frame descriptor record area allocations. Thanks for persevering with this - in my opinion it's now looking like it was worth the effort :) AFAICS the ioremap_wc() that this leads to does appear to give back something non-cacheable on PPC (assuming "pgprot_noncached_wc" isn't horrendously misnamed), and "no-map" should rule out any cacheable linear map alias existing, so it would seem that this approach should avert Scott's concerns about attribute mismatches. > Signed-off-by: Roy Pledge <roy.pledge@nxp.com> > --- > drivers/soc/fsl/qbman/qman_ccsr.c | 119 +++++++++++++++++++++----------------- > drivers/soc/fsl/qbman/qman_priv.h | 4 +- > drivers/soc/fsl/qbman/qman_test.h | 2 - > 3 files changed, 68 insertions(+), 57 deletions(-) > > diff --git a/drivers/soc/fsl/qbman/qman_ccsr.c b/drivers/soc/fsl/qbman/qman_ccsr.c > index 90bc40c..35c59ca 100644 > --- a/drivers/soc/fsl/qbman/qman_ccsr.c > +++ b/drivers/soc/fsl/qbman/qman_ccsr.c > @@ -400,63 +400,19 @@ static int qm_init_pfdr(struct device *dev, u32 pfdr_start, u32 num) > return -ENODEV; > } > > -/* > - * Ideally we would use the DMA API to turn rmem->base into a DMA address > - * (especially if iommu translations ever get involved). Unfortunately, the > - * DMA API currently does not allow mapping anything that is not backed with > - * a struct page. > - */ > +/* QMan needs two global memory areas initialized at boot time: > + 1) FQD: Frame Queue Descriptors used to manage frame queues > + 2) PFDR: Packed Frame Queue Descriptor Records used to store frames > + Both areas are reserved using the device tree reserved memory framework > + and the addresses and sizes are initialized when the QMan device is probed */ > static dma_addr_t fqd_a, pfdr_a; > static size_t fqd_sz, pfdr_sz; > > -static int qman_fqd(struct reserved_mem *rmem) > -{ > - fqd_a = rmem->base; > - fqd_sz = rmem->size; > - > - WARN_ON(!(fqd_a && fqd_sz)); > - > - return 0; > -} > -RESERVEDMEM_OF_DECLARE(qman_fqd, "fsl,qman-fqd", qman_fqd); > - > -static int qman_pfdr(struct reserved_mem *rmem) > -{ > - pfdr_a = rmem->base; > - pfdr_sz = rmem->size; > - > - WARN_ON(!(pfdr_a && pfdr_sz)); > - > - return 0; > -} > -RESERVEDMEM_OF_DECLARE(qman_pfdr, "fsl,qman-pfdr", qman_pfdr); > - I notice that patch #1 isn't removing the equivalent bman_fbpr() handler and declaration as here - is that deliberate or just an oversight? > static unsigned int qm_get_fqid_maxcnt(void) > { > return fqd_sz / 64; > } > > -/* > - * Flush this memory range from data cache so that QMAN originated > - * transactions for this memory region could be marked non-coherent. > - */ > -static int zero_priv_mem(struct device *dev, struct device_node *node, > - phys_addr_t addr, size_t sz) > -{ > - /* map as cacheable, non-guarded */ > - void __iomem *tmpp = ioremap_prot(addr, sz, 0); > - > - if (!tmpp) > - return -ENOMEM; > - > - memset_io(tmpp, 0, sz); > - flush_dcache_range((unsigned long)tmpp, > - (unsigned long)tmpp + sz); > - iounmap(tmpp); > - > - return 0; > -} > - > static void log_edata_bits(struct device *dev, u32 bit_count) > { > u32 i, j, mask = 0xffffffff; > @@ -687,11 +643,12 @@ static int qman_resource_init(struct device *dev) > static int fsl_qman_probe(struct platform_device *pdev) > { > struct device *dev = &pdev->dev; > - struct device_node *node = dev->of_node; > + struct device_node *mem_node, *node = dev->of_node; > struct resource *res; > int ret, err_irq; > u16 id; > u8 major, minor; > + u64 size; > > res = platform_get_resource(pdev, IORESOURCE_MEM, 0); > if (!res) { > @@ -727,10 +684,66 @@ static int fsl_qman_probe(struct platform_device *pdev) > qm_channel_caam = QMAN_CHANNEL_CAAM_REV3; > } > > - ret = zero_priv_mem(dev, node, fqd_a, fqd_sz); > - WARN_ON(ret); > - if (ret) > + /* Order of memory regions is assumed as FQD followed by PFDR > + in order to ensure allocations from the correct regions the > + driver initializes then allocates each piece in order */ > + > + ret = of_reserved_mem_device_init_by_idx(dev, dev->of_node, 0); > + if (ret) { > + dev_err(dev, "of_reserved_mem_device_init_by_idx(0) failed 0x%x\n", > + ret); > + return -ENODEV; > + } > + mem_node = of_parse_phandle(dev->of_node, "memory-region", 0); > + if (mem_node) { > + ret = of_property_read_u64(mem_node, "size", &size); > + if (ret) { > + dev_err(dev, "FQD: of_address_to_resource fails 0x%x\n", ret); Nit: of_address_to_resource? > + return -ENODEV; > + } > + fqd_sz = size; > + } else { > + dev_err(dev, "No memory-region found for FQD\n"); > + return -ENODEV; > + } > + if (!dma_zalloc_coherent(dev, fqd_sz, &fqd_a, 0)) { > + dev_err(dev, "Alloc FQD memory failed\n"); > + return -ENODEV; > + } Between here, below, and patch #1, it looks like this "init, get size, alloc" pattern could be nicely factored out into a common helper function. > + > + dev_info(dev, "Allocated FQD 0x%llx 0x%zx\n", fqd_a, fqd_sz); > + > + /* Disassociate the FQD reseverd memory area from the device because > + a device can only have one DMA memory area. This should be fine > + since the memory is allocated and initalized and only ever accessed > + by the QMan device from now on */ Typos: reserved, initialized. > + of_reserved_mem_device_release(dev); I see the driver does not implement a .remove method where you would otherwise be technically expected to balance the allocation with a dma_free_coherent(), so I think this trick is acceptable. If anyone wants to change that in future, it wouldn't seem unreasonable to extend the core code to handle multiple areas per device (and at worst, I guess you could deviously juggle dev->dma_mem around the DMA API calls). Other than those first few nits, and any possible PPC-specific angles I'm not aware of, I think the series is looking good! Thanks, Robin. > + > + /* Setup PFDR memory */ > + ret = of_reserved_mem_device_init_by_idx(dev, dev->of_node, 1); > + if (ret) { > + dev_err(dev, "of_reserved_mem_device_init(1) failed 0x%x\n", > + ret); > + return -ENODEV; > + } > + mem_node = of_parse_phandle(dev->of_node, "memory-region", 1); > + if (mem_node) { > + ret = of_property_read_u64(mem_node, "size", &size); > + if (ret) { > + dev_err(dev, "PFDR: of_address_to_resource fails 0x%x\n", ret); > + return -ENODEV; > + } > + pfdr_sz = size; > + } else { > + dev_err(dev, "No memory-region found for PFDR\n"); > + return -ENODEV; > + } > + if (!dma_zalloc_coherent(dev, pfdr_sz, &pfdr_a, 0)) { > + dev_err(dev, "Alloc PFDR Failed size 0x%zx\n", pfdr_sz); > return -ENODEV; > + } > + > + dev_info(dev, "Allocated PFDR 0x%llx 0x%zx\n", pfdr_a, pfdr_sz); > > ret = qman_init_ccsr(dev); > if (ret) { > diff --git a/drivers/soc/fsl/qbman/qman_priv.h b/drivers/soc/fsl/qbman/qman_priv.h > index 22725bd..1e998ea5 100644 > --- a/drivers/soc/fsl/qbman/qman_priv.h > +++ b/drivers/soc/fsl/qbman/qman_priv.h > @@ -28,12 +28,12 @@ > * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. > */ > > -#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt > - > #include "dpaa_sys.h" > > #include <soc/fsl/qman.h> > #include <linux/iommu.h> > +#include <linux/dma-contiguous.h> > +#include <linux/of_address.h> > > #if defined(CONFIG_FSL_PAMU) > #include <asm/fsl_pamu_stash.h> > diff --git a/drivers/soc/fsl/qbman/qman_test.h b/drivers/soc/fsl/qbman/qman_test.h > index d5f8cb2..41bdbc48 100644 > --- a/drivers/soc/fsl/qbman/qman_test.h > +++ b/drivers/soc/fsl/qbman/qman_test.h > @@ -30,7 +30,5 @@ > > #include "qman_priv.h" > > -#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt > - > int qman_test_stash(void); > int qman_test_api(void); >
Robin Murphy <robin.murphy@arm.com> writes: > Hi Roy, > > On 29/03/17 22:13, Roy Pledge wrote: >> Use the shared-memory-pool mechanism for frame queue descriptor and >> packed frame descriptor record area allocations. > > Thanks for persevering with this - in my opinion it's now looking like > it was worth the effort :) > > AFAICS the ioremap_wc() that this leads to does appear to give back > something non-cacheable on PPC (assuming "pgprot_noncached_wc" isn't > horrendously misnamed), and "no-map" should rule out any cacheable > linear map alias existing, so it would seem that this approach should > avert Scott's concerns about attribute mismatches. How does 'no-map' translate into something being excluded from the linear mapping? cheers
On 31/03/17 04:27, Michael Ellerman wrote: > Robin Murphy <robin.murphy@arm.com> writes: > >> Hi Roy, >> >> On 29/03/17 22:13, Roy Pledge wrote: >>> Use the shared-memory-pool mechanism for frame queue descriptor and >>> packed frame descriptor record area allocations. >> >> Thanks for persevering with this - in my opinion it's now looking like >> it was worth the effort :) >> >> AFAICS the ioremap_wc() that this leads to does appear to give back >> something non-cacheable on PPC (assuming "pgprot_noncached_wc" isn't >> horrendously misnamed), and "no-map" should rule out any cacheable >> linear map alias existing, so it would seem that this approach should >> avert Scott's concerns about attribute mismatches. > > How does 'no-map' translate into something being excluded from the > linear mapping? Reserved regions marked with "no-map" get memblock_remove()d by early_init_dt_alloc_reserved_memory_arch(). As I understand things, the linear map should only cover memblock areas, and it would be explicitly violating the semantics of "no-map" to still cover such a region. Robin. > > cheers >
On Fri, 2017-03-31 at 18:55 +0100, Robin Murphy wrote: > On 31/03/17 04:27, Michael Ellerman wrote: > > > > Robin Murphy <robin.murphy@arm.com> writes: > > > > > > > > Hi Roy, > > > > > > On 29/03/17 22:13, Roy Pledge wrote: > > > > > > > > Use the shared-memory-pool mechanism for frame queue descriptor and > > > > packed frame descriptor record area allocations. > > > Thanks for persevering with this - in my opinion it's now looking like > > > it was worth the effort :) > > > > > > AFAICS the ioremap_wc() that this leads to does appear to give back > > > something non-cacheable on PPC (assuming "pgprot_noncached_wc" isn't > > > horrendously misnamed), and "no-map" should rule out any cacheable > > > linear map alias existing, so it would seem that this approach should > > > avert Scott's concerns about attribute mismatches. > > How does 'no-map' translate into something being excluded from the > > linear mapping? > Reserved regions marked with "no-map" get memblock_remove()d by > early_init_dt_alloc_reserved_memory_arch(). As I understand things, the > linear map should only cover memblock areas, and it would be explicitly > violating the semantics of "no-map" to still cover such a region. Discontiguous memory isn't supported on these PPC chips. Everything up to memblock_end_of_DRAM() gets mapped -- and if that were to change, the fragmentation would waste TLB1 entries. This also breaks compatibility with existing device trees. I suggest putting an ifdef in the qbman driver to add the new scheme for non-PPC arches only. -Scott
On 01/04/17 08:25, Scott Wood wrote: > On Fri, 2017-03-31 at 18:55 +0100, Robin Murphy wrote: >> On 31/03/17 04:27, Michael Ellerman wrote: >>> >>> Robin Murphy <robin.murphy@arm.com> writes: >>> >>>> >>>> Hi Roy, >>>> >>>> On 29/03/17 22:13, Roy Pledge wrote: >>>>> >>>>> Use the shared-memory-pool mechanism for frame queue descriptor and >>>>> packed frame descriptor record area allocations. >>>> Thanks for persevering with this - in my opinion it's now looking like >>>> it was worth the effort :) >>>> >>>> AFAICS the ioremap_wc() that this leads to does appear to give back >>>> something non-cacheable on PPC (assuming "pgprot_noncached_wc" isn't >>>> horrendously misnamed), and "no-map" should rule out any cacheable >>>> linear map alias existing, so it would seem that this approach should >>>> avert Scott's concerns about attribute mismatches. >>> How does 'no-map' translate into something being excluded from the >>> linear mapping? >> Reserved regions marked with "no-map" get memblock_remove()d by >> early_init_dt_alloc_reserved_memory_arch(). As I understand things, the >> linear map should only cover memblock areas, and it would be explicitly >> violating the semantics of "no-map" to still cover such a region. > > Discontiguous memory isn't supported on these PPC chips. Everything up to > memblock_end_of_DRAM() gets mapped -- and if that were to change, the > fragmentation would waste TLB1 entries. Ah, so the "PPC-specific angles I'm not aware of" category is indeed non-empty - I guess the lack of HAVE_GENERIC_DMA_COHERENT might be related, then. That said, though, AFAICS only certain x86 and s390 configurations ever call memblock_set_bottom_up(true), so we should be able to assume that the reserved region allocations always fall through to __memblock_find_range_top_down(). Thus if your DRAM is contiguous, then "no-map"-ing the reserved regions will simply end up pushing memblock_end_of_DRAM() down in a manner that would appear to still avoid overlaps. I can only see that going wrong if the end of DRAM wasn't at least 32MB aligned to begin with - is that ever likely to happen in practice? Robin. > This also breaks compatibility with existing device trees. I suggest putting > an ifdef in the qbman driver to add the new scheme for non-PPC arches only. > > -Scott >
On Mon, 2017-04-03 at 15:52 +0100, Robin Murphy wrote: > On 01/04/17 08:25, Scott Wood wrote: > > > > On Fri, 2017-03-31 at 18:55 +0100, Robin Murphy wrote: > > > > > > On 31/03/17 04:27, Michael Ellerman wrote: > > > > > > > > > > > > Robin Murphy <robin.murphy@arm.com> writes: > > > > > > > > > > > > > > > > > > > Hi Roy, > > > > > > > > > > On 29/03/17 22:13, Roy Pledge wrote: > > > > > > > > > > > > > > > > > > Use the shared-memory-pool mechanism for frame queue descriptor > > > > > > and > > > > > > packed frame descriptor record area allocations. > > > > > Thanks for persevering with this - in my opinion it's now looking > > > > > like > > > > > it was worth the effort :) > > > > > > > > > > AFAICS the ioremap_wc() that this leads to does appear to give back > > > > > something non-cacheable on PPC (assuming "pgprot_noncached_wc" isn't > > > > > horrendously misnamed), and "no-map" should rule out any cacheable > > > > > linear map alias existing, so it would seem that this approach > > > > > should > > > > > avert Scott's concerns about attribute mismatches. > > > > How does 'no-map' translate into something being excluded from the > > > > linear mapping? > > > Reserved regions marked with "no-map" get memblock_remove()d by > > > early_init_dt_alloc_reserved_memory_arch(). As I understand things, the > > > linear map should only cover memblock areas, and it would be explicitly > > > violating the semantics of "no-map" to still cover such a region. > > Discontiguous memory isn't supported on these PPC chips. Everything up to > > memblock_end_of_DRAM() gets mapped -- and if that were to change, the > > fragmentation would waste TLB1 entries. > Ah, so the "PPC-specific angles I'm not aware of" category is indeed > non-empty - I guess the lack of HAVE_GENERIC_DMA_COHERENT might be > related, then. > > That said, though, AFAICS only certain x86 and s390 configurations ever > call memblock_set_bottom_up(true), so we should be able to assume that > the reserved region allocations always fall through to > __memblock_find_range_top_down(). Thus if your DRAM is contiguous, then > "no-map"-ing the reserved regions will simply end up pushing > memblock_end_of_DRAM() down in a manner that would appear to still avoid > overlaps. Can you guarantee it will be at the end? What if there were other early memblock allocations (e.g. other reserved-memory regions without no-map) that came first? > I can only see that going wrong if the end of DRAM wasn't at > least 32MB aligned to begin with - is that ever likely to happen in > practice? Probably not, unless the user asks for an unusual memory size via mem= or similar. -Scott
diff --git a/drivers/soc/fsl/qbman/qman_ccsr.c b/drivers/soc/fsl/qbman/qman_ccsr.c index 90bc40c..35c59ca 100644 --- a/drivers/soc/fsl/qbman/qman_ccsr.c +++ b/drivers/soc/fsl/qbman/qman_ccsr.c @@ -400,63 +400,19 @@ static int qm_init_pfdr(struct device *dev, u32 pfdr_start, u32 num) return -ENODEV; } -/* - * Ideally we would use the DMA API to turn rmem->base into a DMA address - * (especially if iommu translations ever get involved). Unfortunately, the - * DMA API currently does not allow mapping anything that is not backed with - * a struct page. - */ +/* QMan needs two global memory areas initialized at boot time: + 1) FQD: Frame Queue Descriptors used to manage frame queues + 2) PFDR: Packed Frame Queue Descriptor Records used to store frames + Both areas are reserved using the device tree reserved memory framework + and the addresses and sizes are initialized when the QMan device is probed */ static dma_addr_t fqd_a, pfdr_a; static size_t fqd_sz, pfdr_sz; -static int qman_fqd(struct reserved_mem *rmem) -{ - fqd_a = rmem->base; - fqd_sz = rmem->size; - - WARN_ON(!(fqd_a && fqd_sz)); - - return 0; -} -RESERVEDMEM_OF_DECLARE(qman_fqd, "fsl,qman-fqd", qman_fqd); - -static int qman_pfdr(struct reserved_mem *rmem) -{ - pfdr_a = rmem->base; - pfdr_sz = rmem->size; - - WARN_ON(!(pfdr_a && pfdr_sz)); - - return 0; -} -RESERVEDMEM_OF_DECLARE(qman_pfdr, "fsl,qman-pfdr", qman_pfdr); - static unsigned int qm_get_fqid_maxcnt(void) { return fqd_sz / 64; } -/* - * Flush this memory range from data cache so that QMAN originated - * transactions for this memory region could be marked non-coherent. - */ -static int zero_priv_mem(struct device *dev, struct device_node *node, - phys_addr_t addr, size_t sz) -{ - /* map as cacheable, non-guarded */ - void __iomem *tmpp = ioremap_prot(addr, sz, 0); - - if (!tmpp) - return -ENOMEM; - - memset_io(tmpp, 0, sz); - flush_dcache_range((unsigned long)tmpp, - (unsigned long)tmpp + sz); - iounmap(tmpp); - - return 0; -} - static void log_edata_bits(struct device *dev, u32 bit_count) { u32 i, j, mask = 0xffffffff; @@ -687,11 +643,12 @@ static int qman_resource_init(struct device *dev) static int fsl_qman_probe(struct platform_device *pdev) { struct device *dev = &pdev->dev; - struct device_node *node = dev->of_node; + struct device_node *mem_node, *node = dev->of_node; struct resource *res; int ret, err_irq; u16 id; u8 major, minor; + u64 size; res = platform_get_resource(pdev, IORESOURCE_MEM, 0); if (!res) { @@ -727,10 +684,66 @@ static int fsl_qman_probe(struct platform_device *pdev) qm_channel_caam = QMAN_CHANNEL_CAAM_REV3; } - ret = zero_priv_mem(dev, node, fqd_a, fqd_sz); - WARN_ON(ret); - if (ret) + /* Order of memory regions is assumed as FQD followed by PFDR + in order to ensure allocations from the correct regions the + driver initializes then allocates each piece in order */ + + ret = of_reserved_mem_device_init_by_idx(dev, dev->of_node, 0); + if (ret) { + dev_err(dev, "of_reserved_mem_device_init_by_idx(0) failed 0x%x\n", + ret); + return -ENODEV; + } + mem_node = of_parse_phandle(dev->of_node, "memory-region", 0); + if (mem_node) { + ret = of_property_read_u64(mem_node, "size", &size); + if (ret) { + dev_err(dev, "FQD: of_address_to_resource fails 0x%x\n", ret); + return -ENODEV; + } + fqd_sz = size; + } else { + dev_err(dev, "No memory-region found for FQD\n"); + return -ENODEV; + } + if (!dma_zalloc_coherent(dev, fqd_sz, &fqd_a, 0)) { + dev_err(dev, "Alloc FQD memory failed\n"); + return -ENODEV; + } + + dev_info(dev, "Allocated FQD 0x%llx 0x%zx\n", fqd_a, fqd_sz); + + /* Disassociate the FQD reseverd memory area from the device because + a device can only have one DMA memory area. This should be fine + since the memory is allocated and initalized and only ever accessed + by the QMan device from now on */ + of_reserved_mem_device_release(dev); + + /* Setup PFDR memory */ + ret = of_reserved_mem_device_init_by_idx(dev, dev->of_node, 1); + if (ret) { + dev_err(dev, "of_reserved_mem_device_init(1) failed 0x%x\n", + ret); + return -ENODEV; + } + mem_node = of_parse_phandle(dev->of_node, "memory-region", 1); + if (mem_node) { + ret = of_property_read_u64(mem_node, "size", &size); + if (ret) { + dev_err(dev, "PFDR: of_address_to_resource fails 0x%x\n", ret); + return -ENODEV; + } + pfdr_sz = size; + } else { + dev_err(dev, "No memory-region found for PFDR\n"); + return -ENODEV; + } + if (!dma_zalloc_coherent(dev, pfdr_sz, &pfdr_a, 0)) { + dev_err(dev, "Alloc PFDR Failed size 0x%zx\n", pfdr_sz); return -ENODEV; + } + + dev_info(dev, "Allocated PFDR 0x%llx 0x%zx\n", pfdr_a, pfdr_sz); ret = qman_init_ccsr(dev); if (ret) { diff --git a/drivers/soc/fsl/qbman/qman_priv.h b/drivers/soc/fsl/qbman/qman_priv.h index 22725bd..1e998ea5 100644 --- a/drivers/soc/fsl/qbman/qman_priv.h +++ b/drivers/soc/fsl/qbman/qman_priv.h @@ -28,12 +28,12 @@ * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */ -#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt - #include "dpaa_sys.h" #include <soc/fsl/qman.h> #include <linux/iommu.h> +#include <linux/dma-contiguous.h> +#include <linux/of_address.h> #if defined(CONFIG_FSL_PAMU) #include <asm/fsl_pamu_stash.h> diff --git a/drivers/soc/fsl/qbman/qman_test.h b/drivers/soc/fsl/qbman/qman_test.h index d5f8cb2..41bdbc48 100644 --- a/drivers/soc/fsl/qbman/qman_test.h +++ b/drivers/soc/fsl/qbman/qman_test.h @@ -30,7 +30,5 @@ #include "qman_priv.h" -#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt - int qman_test_stash(void); int qman_test_api(void);
Use the shared-memory-pool mechanism for frame queue descriptor and packed frame descriptor record area allocations. Signed-off-by: Roy Pledge <roy.pledge@nxp.com> --- drivers/soc/fsl/qbman/qman_ccsr.c | 119 +++++++++++++++++++++----------------- drivers/soc/fsl/qbman/qman_priv.h | 4 +- drivers/soc/fsl/qbman/qman_test.h | 2 - 3 files changed, 68 insertions(+), 57 deletions(-)