fix SG list handling in the sparc32 iommu driver
Hi Dave,
this series fixes some long standing bugs in the sparc32 iommu driver, mostly the lack of handling of large sglist offsets in the map_sg method, but also a few other smaller bits. These now show up all the time do some block layer changes in linux-next.
PATCH 1/9 sparc/iommu: use !PageHighMem to check if a page has a kernel mapping
This deobsfucates the check a bit, and prepares for future changes.
--- arch/sparc/mm/iommu.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
arch/sparc/mm/iommu.c +2 -1
@@ -273,7 +273,8 @@ static int sbus_iommu_map_sg_pflush(struct device *dev, struct scatterlist *sgl,* XXX Is this a good assumption? * XXX What if someone else unmaps it here and races us? */ -if ((page = (unsigned long )page_address (sg_page (sg))) !=0 ) { +if (!PageHighMem (sg_page (sg))) { + page = (unsigned long )page_address (sg_page (sg));for (i =0 ; i < n; i++) {if (page != oldpage) {/* Already flushed? */ flush_page_for_dma (page);
Signature · 1 line
2.20.1
PATCH 2/9 sparc/iommu: use sbus_iommu_unmap_page in sbus_iommu_unmap_sg
In reply to Christoph HellwigUse the page-level helper instead of duplicating the logic, while also fixing the incorrect handling of larger than page sized offsets in the sg variant.
--- arch/sparc/mm/iommu.c | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)
arch/sparc/mm/iommu.c +3 -3
@@ -321,11 +321,11 @@ static void sbus_iommu_unmap_sg(struct device *dev, struct scatterlist *sgl, int nents, enum dma_data_direction dir, unsigned long attrs) { struct scatterlist *sg; -int i, n; +int i;for_each_sg (sgl, sg, nents, i) { - n = (sg->length + sg->offset +PAGE_SIZE -1 ) >>PAGE_SHIFT ; -iommu_release_one (dev, sg->dma_address &PAGE_MASK , n); +sbus_iommu_unmap_page (dev, sg->dma_address, sg->length, dir, + attrs); sg->dma_address =0x21212121 ; } }
Signature · 1 line
2.20.1
PATCH 3/9 sparc/iommu: merge iommu_release_one and sbus_iommu_unmap_page
In reply to Christoph HellwigThere is only one caller of iommu_release_one left, so merge it into that one to clean things up a bit.
--- arch/sparc/mm/iommu.c | 21 +++++++-------------- 1 file changed, 7 insertions(+), 14 deletions(-)
arch/sparc/mm/iommu.c +7 -14
@@ -291,14 +291,17 @@ static int sbus_iommu_map_sg_pflush(struct device *dev, struct scatterlist *sgl,return nents; } -static void iommu_release_one (struct device *dev,u32 busa,int npages) +static void sbus_iommu_unmap_page (struct device *dev,dma_addr_t dma_addr, +size_t len,enum dma_data_direction dir,unsigned long attrs) {struct iommu_struct *iommu = dev->archdata.iommu; -int ioptex; -int i; +unsigned int busa = dma_addr &PAGE_MASK ; +unsigned long off = dma_addr & ~PAGE_MASK ; +unsigned int npages = (off + len +PAGE_SIZE -1 ) >>PAGE_SHIFT ; +unsigned int ioptex = (busa - iommu->start) >>PAGE_SHIFT ; +unsigned int i;BUG_ON (busa < iommu->start); - ioptex = (busa - iommu->start) >>PAGE_SHIFT ;for (i =0 ; i < npages; i++) {iopte_val (iommu->page_table[ioptex + i]) =0 ;iommu_invalidate_page (iommu->regs, busa);
@@ -307,16 +310,6 @@ static void iommu_release_one(struct device *dev, u32 busa, int npages)bit_map_clear (&iommu->usemap, ioptex, npages); } -static void sbus_iommu_unmap_page (struct device *dev,dma_addr_t dma_addr, -size_t len,enum dma_data_direction dir,unsigned long attrs) -{ -unsigned long off = dma_addr & ~PAGE_MASK ; -int npages; - - npages = (off + len +PAGE_SIZE -1 ) >>PAGE_SHIFT ; -iommu_release_one (dev, dma_addr &PAGE_MASK , npages); -} -static void sbus_iommu_unmap_sg (struct device *dev,struct scatterlist *sgl,int nents,enum dma_data_direction dir,unsigned long attrs) {
Signature · 1 line
2.20.1
PATCH 4/9 sparc/iommu: create a common helper for map_sg
In reply to Christoph HellwigShare the code for the global and per-page flush map_sg loops using a simple bool parameter to disable the per-page flush for the former variant.
--- arch/sparc/mm/iommu.c | 37 +++++++++++++++++-------------------- 1 file changed, 17 insertions(+), 20 deletions(-)
arch/sparc/mm/iommu.c +17 -20
@@ -241,25 +241,9 @@ static dma_addr_t sbus_iommu_map_page_pflush(struct device *dev,return __sbus_iommu_map_page (dev, page, offset, len); } -static int sbus_iommu_map_sg_gflush (struct device *dev,struct scatterlist *sgl, -int nents,enum dma_data_direction dir,unsigned long attrs) -{ -struct scatterlist *sg; -int i, n; - -flush_page_for_dma (0 ); - -for_each_sg (sgl, sg, nents, i) { - n = (sg->length + sg->offset +PAGE_SIZE -1 ) >>PAGE_SHIFT ; - sg->dma_address =iommu_get_one (dev,sg_page (sg), n) + sg->offset; - sg->dma_length = sg->length; - } - -return nents; -} - -static int sbus_iommu_map_sg_pflush (struct device *dev,struct scatterlist *sgl, -int nents,enum dma_data_direction dir,unsigned long attrs) +static int __sbus_iommu_map_sg (struct device *dev,struct scatterlist *sgl, +int nents,enum dma_data_direction dir,unsigned long attrs, +bool per_page_flush) {unsigned long page, oldpage =0 ;struct scatterlist *sg;
@@ -273,7 +257,7 @@ static int sbus_iommu_map_sg_pflush(struct device *dev, struct scatterlist *sgl,* XXX Is this a good assumption? * XXX What if someone else unmaps it here and races us? */ -if (!PageHighMem (sg_page (sg))) { +if (per_page_flush && !PageHighMem (sg_page (sg))) { page = (unsigned long )page_address (sg_page (sg));for (i =0 ; i < n; i++) {if (page != oldpage) {/* Already flushed? */
@@ -291,6 +275,19 @@ static int sbus_iommu_map_sg_pflush(struct device *dev, struct scatterlist *sgl,return nents; } +static int sbus_iommu_map_sg_gflush (struct device *dev,struct scatterlist *sgl, +int nents,enum dma_data_direction dir,unsigned long attrs) +{ +flush_page_for_dma (0 ); +return __sbus_iommu_map_sg (dev, sgl, nents, dir, attrs, false); +} + +static int sbus_iommu_map_sg_pflush (struct device *dev,struct scatterlist *sgl, +int nents,enum dma_data_direction dir,unsigned long attrs) +{ +return __sbus_iommu_map_sg (dev, sgl, nents, dir, attrs, true); +} +static void sbus_iommu_unmap_page (struct device *dev,dma_addr_t dma_addr,size_t len,enum dma_data_direction dir,unsigned long attrs) {
Signature · 1 line
2.20.1
PATCH 5/9 sparc/iommu: pass a physical address to iommu_get_one
In reply to Christoph HellwigNo need for the page structure, just the paddr / pfn. This is going to simplify fixes to the callers.
--- arch/sparc/mm/iommu.c | 13 +++++++------ 1 file changed, 7 insertions(+), 6 deletions(-)
arch/sparc/mm/iommu.c +7 -6
@@ -175,16 +175,17 @@ static void iommu_flush_iotlb(iopte_t *iopte, unsigned int niopte) } } -static u32 iommu_get_one (struct device *dev,struct page *page,int npages) +static u32 iommu_get_one (struct device *dev,phys_addr_t paddr,int npages) {struct iommu_struct *iommu = dev->archdata.iommu;int ioptex;iopte_t *iopte, *iopte0;unsigned int busa, busa0; +unsigned long pfn =__phys_to_pfn (paddr);int i;/* page color = pfn of page */ - ioptex =bit_map_string_get (&iommu->usemap, npages,page_to_pfn (page)); + ioptex =bit_map_string_get (&iommu->usemap, npages, pfn);if (ioptex <0 )panic ("iommu out" ); busa0 = iommu->start + (ioptex <<PAGE_SHIFT );
@@ -193,11 +194,11 @@ static u32 iommu_get_one(struct device *dev, struct page *page, int npages) busa = busa0; iopte = iopte0;for (i =0 ; i < npages; i++) { -iopte_val (*iopte) =MKIOPTE (page_to_pfn (page),IOPERM ); +iopte_val (*iopte) =MKIOPTE (pfn,IOPERM );iommu_invalidate_page (iommu->regs, busa); busa +=PAGE_SIZE ; iopte++; - page++; + pfn++; }iommu_flush_iotlb (iopte0, npages);
@@ -215,7 +216,7 @@ static dma_addr_t __sbus_iommu_map_page(struct device *dev, struct page *page,/* XXX So what is maxphys for us and how do drivers know it? */ if (!len || len >256 *1024 )return DMA_MAPPING_ERROR ; -return iommu_get_one (dev,virt_to_page (vaddr), npages) + off; +return iommu_get_one (dev,virt_to_phys (vaddr), npages) + off; } static dma_addr_t sbus_iommu_map_page_gflush(struct device *dev,
@@ -268,7 +269,7 @@ static int __sbus_iommu_map_sg(struct device *dev, struct scatterlist *sgl, } } - sg->dma_address =iommu_get_one (dev,sg_page (sg), n) + sg->offset; + sg->dma_address =iommu_get_one (dev,sg_phys (sg), n) + sg->offset; sg->dma_length = sg->length; }
Signature · 1 line
2.20.1
PATCH 6/9 sparc/iommu: move per-page flushing into __sbus_iommu_map_page
In reply to Christoph HellwigThis prepares for reusing __sbus_iommu_map_page in the map_sg path.
--- arch/sparc/mm/iommu.c | 26 ++++++++++++++------------ 1 file changed, 14 insertions(+), 12 deletions(-)
arch/sparc/mm/iommu.c +14 -12
@@ -207,15 +207,25 @@ static u32 iommu_get_one(struct device *dev, phys_addr_t paddr, int npages) }static dma_addr_t __sbus_iommu_map_page (struct device *dev,struct page *page, -unsigned long offset,size_t len) +unsigned long offset,size_t len,bool per_page_flush) {void *vaddr =page_address (page) + offset;unsigned long off = (unsigned long )vaddr & ~PAGE_MASK ;unsigned long npages = (off + len +PAGE_SIZE -1 ) >>PAGE_SHIFT ; - +/* XXX So what is maxphys for us and how do drivers know it? */ if (!len || len >256 *1024 )return DMA_MAPPING_ERROR ; + +if (per_page_flush) { +unsigned long p = (unsigned long )vaddr &PAGE_MASK ; + +while (p < (unsigned long )vaddr + len) { +flush_page_for_dma (p); + p +=PAGE_SIZE ; + } + } +return iommu_get_one (dev,virt_to_phys (vaddr), npages) + off; }
@@ -224,22 +234,14 @@ static dma_addr_t sbus_iommu_map_page_gflush(struct device *dev, enum dma_data_direction dir, unsigned long attrs) {flush_page_for_dma (0 ); -return __sbus_iommu_map_page (dev, page, offset, len); +return __sbus_iommu_map_page (dev, page, offset, len, false); }static dma_addr_t sbus_iommu_map_page_pflush (struct device *dev,struct page *page,unsigned long offset,size_t len,enum dma_data_direction dir,unsigned long attrs) { -void *vaddr =page_address (page) + offset; -unsigned long p = ((unsigned long )vaddr) &PAGE_MASK ; - -while (p < (unsigned long )vaddr + len) { -flush_page_for_dma (p); - p +=PAGE_SIZE ; - } - -return __sbus_iommu_map_page (dev, page, offset, len); +return __sbus_iommu_map_page (dev, page, offset, len, true); } static int __sbus_iommu_map_sg(struct device *dev, struct scatterlist *sgl,
Signature · 1 line
2.20.1
PATCH 7/9 sparc/iommu: fix __sbus_iommu_map_page for highmem pages
In reply to Christoph Hellwig__sbus_iommu_map_page currently assumes all pages are mapped into the kernel direct mapping. Switch to using physical address instead of virtual ones for all the normal mapping operations, and only use the virtual addresses for cache flushing when not operating on a highmem page.
--- arch/sparc/mm/iommu.c | 15 +++++++-------- 1 file changed, 7 insertions(+), 8 deletions(-)
arch/sparc/mm/iommu.c +7 -8
@@ -209,24 +209,23 @@ static u32 iommu_get_one(struct device *dev, phys_addr_t paddr, int npages)static dma_addr_t __sbus_iommu_map_page (struct device *dev,struct page *page,unsigned long offset,size_t len,bool per_page_flush) { -void *vaddr =page_address (page) + offset; -unsigned long off = (unsigned long )vaddr & ~PAGE_MASK ; +phys_addr_t paddr =page_to_phys (page) + offset; +unsigned long off = paddr & ~PAGE_MASK ;unsigned long npages = (off + len +PAGE_SIZE -1 ) >>PAGE_SHIFT ;/* XXX So what is maxphys for us and how do drivers know it? */ if (!len || len >256 *1024 )return DMA_MAPPING_ERROR ; -if (per_page_flush) { -unsigned long p = (unsigned long )vaddr &PAGE_MASK ; +if (per_page_flush && !PageHighMem (page)) { +unsigned long vaddr, p; -while (p < (unsigned long )vaddr + len) { + vaddr = (unsigned long )page_address (page) + offset; +for (p = vaddr &PAGE_MASK ; p < vaddr + len; p +=PAGE_SIZE )flush_page_for_dma (p); - p +=PAGE_SIZE ; - } } -return iommu_get_one (dev,virt_to_phys (vaddr), npages) + off; +return iommu_get_one (dev, paddr, npages) + off; } static dma_addr_t sbus_iommu_map_page_gflush(struct device *dev,
Signature · 1 line
2.20.1
PATCH 8/9 sparc/iommu: use __sbus_iommu_map_page to implement the map_sg path
In reply to Christoph HellwigThis means we handle > PAGE_SIZE offsets fine, and grow the size check so far only performed in the map_page path. We lose the optimization to not double flush a page if it apears in multiple consecutive SG list entries. But at least for block I/O those don't happen anymore since we properly merge in higher layers anyway.
--- arch/sparc/mm/iommu.c | 31 ++++++++++--------------------- 1 file changed, 10 insertions(+), 21 deletions(-)
arch/sparc/mm/iommu.c +10 -21
@@ -217,6 +217,11 @@ static dma_addr_t __sbus_iommu_map_page(struct device *dev, struct page *page,if (!len || len >256 *1024 )return DMA_MAPPING_ERROR ; +/* +* We expect unmapped highmem pages to be not in the cache. +* XXX Is this a good assumption? +* XXX What if someone else unmaps it here and races us? +*/ if (per_page_flush && !PageHighMem (page)) {unsigned long vaddr, p;
@@ -247,30 +252,14 @@ static int __sbus_iommu_map_sg(struct device *dev, struct scatterlist *sgl, int nents, enum dma_data_direction dir, unsigned long attrs, bool per_page_flush) { - unsigned long page, oldpage = 0; struct scatterlist *sg; -int i, j, n; +int j;for_each_sg (sgl, sg, nents, j) { - n = (sg->length + sg->offset +PAGE_SIZE -1 ) >>PAGE_SHIFT ; - -/* -* We expect unmapped highmem pages to be not in the cache. -* XXX Is this a good assumption? -* XXX What if someone else unmaps it here and races us? -*/ -if (per_page_flush && !PageHighMem (sg_page (sg))) { - page = (unsigned long )page_address (sg_page (sg)); -for (i =0 ; i < n; i++) { -if (page != oldpage) {/* Already flushed? */ -flush_page_for_dma (page); - oldpage = page; - } - page +=PAGE_SIZE ; - } - } - - sg->dma_address =iommu_get_one (dev,sg_phys (sg), n) + sg->offset; + sg->dma_address =__sbus_iommu_map_page (dev,sg_page (sg), + sg->offset, sg->length, per_page_flush); +if (sg->dma_address =DMA_MAPPING_ERROR ) +return 0 ; sg->dma_length = sg->length; }
Signature · 1 line
2.20.1
PATCH 9/9 sparc/iommu: merge iommu_get_one and __sbus_iommu_map_page
In reply to Christoph HellwigThere is only one caller of iommu_get_one left, so merge it into that one to clean things up a bit.
--- arch/sparc/mm/iommu.c | 56 +++++++++++++++++++------------------------ 1 file changed, 24 insertions(+), 32 deletions(-)
arch/sparc/mm/iommu.c +24 -32
@@ -175,43 +175,17 @@ static void iommu_flush_iotlb(iopte_t *iopte, unsigned int niopte) } } -static u32 iommu_get_one (struct device *dev,phys_addr_t paddr,int npages) -{ -struct iommu_struct *iommu = dev->archdata.iommu; -int ioptex; -iopte_t *iopte, *iopte0; -unsigned int busa, busa0; -unsigned long pfn =__phys_to_pfn (paddr); -int i; - -/* page color = pfn of page */ - ioptex =bit_map_string_get (&iommu->usemap, npages, pfn); -if (ioptex <0 ) -panic ("iommu out" ); - busa0 = iommu->start + (ioptex <<PAGE_SHIFT ); - iopte0 = &iommu->page_table[ioptex]; - - busa = busa0; - iopte = iopte0; -for (i =0 ; i < npages; i++) { -iopte_val (*iopte) =MKIOPTE (pfn,IOPERM ); -iommu_invalidate_page (iommu->regs, busa); - busa +=PAGE_SIZE ; - iopte++; - pfn++; - } - -iommu_flush_iotlb (iopte0, npages); - -return busa0; -} -static dma_addr_t __sbus_iommu_map_page (struct device *dev,struct page *page,unsigned long offset,size_t len,bool per_page_flush) { +struct iommu_struct *iommu = dev->archdata.iommu;phys_addr_t paddr =page_to_phys (page) + offset;unsigned long off = paddr & ~PAGE_MASK ;unsigned long npages = (off + len +PAGE_SIZE -1 ) >>PAGE_SHIFT ; +unsigned long pfn =__phys_to_pfn (paddr); +unsigned int busa, busa0; +iopte_t *iopte, *iopte0; +int ioptex, i;/* XXX So what is maxphys for us and how do drivers know it? */ if (!len || len >256 *1024 )
@@ -230,7 +204,25 @@ static dma_addr_t __sbus_iommu_map_page(struct device *dev, struct page *page,flush_page_for_dma (p); } -return iommu_get_one (dev, paddr, npages) + off; +/* page color = pfn of page */ + ioptex =bit_map_string_get (&iommu->usemap, npages, pfn); +if (ioptex <0 ) +panic ("iommu out" ); + busa0 = iommu->start + (ioptex <<PAGE_SHIFT ); + iopte0 = &iommu->page_table[ioptex]; + + busa = busa0; + iopte = iopte0; +for (i =0 ; i < npages; i++) { +iopte_val (*iopte) =MKIOPTE (pfn,IOPERM ); +iommu_invalidate_page (iommu->regs, busa); + busa +=PAGE_SIZE ; + iopte++; + pfn++; + } + +iommu_flush_iotlb (iopte0, npages); +return busa0 + off; } static dma_addr_t sbus_iommu_map_page_gflush(struct device *dev,
Signature · 1 line
2.20.1
Guenter Roeck
Tested-byQuoted 6 lines · Christoph Hellwig
Hi Dave,
this series fixes some long standing bugs in the sparc32 iommu driver, mostly the lack of handling of large sglist offsets in the
map_sgmethod, but also a few other smaller bits. These now show up all the time do some block layer changes in linux-next.
For the series on top of next-20190416:
... meaning the sparc boot tests no longer crash with the series applied.
Guenter
Hi Dave,
can you pick this series up?
David Miller
maintainerFrom: Christoph Hellwig <[email protected]>
Date: Tue, 16 Apr 2019 20:23:38 +0200
Quoted 4 lines
this series fixes some long standing bugs in the sparc32 iommu driver, mostly the lack of handling of large sglist offsets in the
map_sgmethod, but also a few other smaller bits. These now show up all the time do some block layer changes in linux-next.
Series applied, thanks.