From 64908ad95c34f25849412d6d4735ac10f8fb6575 Mon Sep 17 00:00:00 2001 From: Grant Grundler Date: Fri, 21 Oct 2005 22:37:20 -0400 Subject: [PARISC] Update sba_iommu from parisc tree revert use of %%sr0 in fdc asm. Thanks to Joel Soete for pointing out this oversight. Signed-off-by: Grant Grundler 2.6.14-rc2-pa3 move "sync" outside the main loop that fills IO Pdir. Signed-off-by: Grant Grundler remove explicit use of sr0 in fdc ops. Thanks to Joel Soete for reminding me were I added those... Signed-off-by: Grant Grundler 2.6.14-rc2-pa2 - make SBA more anal about invalidating pdir entries Previous code cleared the valid flag a pdir entry but it did NOT guarantee this change was visible to the PDIR before writing the PCOM register. Ie the SBA could pick up a stale entry if the write happened to hit the SBA before the cacheline was flushed from the cache. Long term, I think I want to make this a compile time flag. Developement tree should enable anal pdir checking by default and Debian can disable it with either a CONFIG option or one-line patch. fdc/sync options can only negatively affect performance though I haven't measure how much yet. If someone can run netperf TCP_RR across gige and compare -pa1 and -pa2, that would be sufficient. Cleaned up the use of "fdc" to make sure it's using "kernel" space id (specify sr0 but maps to sr4-7). It seems a bit fragile to assume "sr1" gets loaded with KERNEL_SPACE which is how the code works today. Tested on 32 and 64-bit SMP kernels on j6k. Signed-off-by: Grant Grundler remove PDC_NARROW from SBA and document history of PDC_NARROW a bit. It will still show up in an older kernel's .config file. Signed-off-by: Grant Grundler if/ifdef cleanups from Joel Soete. Signed-off-by: Grant Grundler 2.6.12-rc4-pa2 fix 32-bit support for Astro platforms o Since my last SBA code change, SBA could allocate more than 1GB of IOVA space on Astro boxes with more than 1GB of RAM when running 32-bit kernel. This is bad since IOMMU can only talk to the first 1GB at most. Kudos to jejb for quickly spotting that bug. o jejb also noted SBA should *always* reject DMA masks > 32-bits since DMA-mapping.txt indicates caller should try again with 32-bits. o off-by-one error when comparing the mask to IOVA space size. Signed-off-by: Grant Grundler Signed-off-by: Kyle McMartin --- drivers/parisc/sba_iommu.c | 140 +++++++++++++++++++++++++++++---------------- 1 file changed, 91 insertions(+), 49 deletions(-) (limited to 'drivers/parisc') diff --git a/drivers/parisc/sba_iommu.c b/drivers/parisc/sba_iommu.c index 6256ad365d0b..48591badd4ce 100644 --- a/drivers/parisc/sba_iommu.c +++ b/drivers/parisc/sba_iommu.c @@ -91,8 +91,8 @@ extern struct proc_dir_entry * proc_mckinley_root; #define DBG_RES(x...) #endif -#if defined(__LP64__) && !defined(CONFIG_PDC_NARROW) -/* "low end" PA8800 machines use ZX1 chipset */ +#if defined(CONFIG_64BIT) +/* "low end" PA8800 machines use ZX1 chipset: PAT PDC and only run 64-bit */ #define ZX1_SUPPORT #endif @@ -231,7 +231,7 @@ struct ioc { spinlock_t res_lock; unsigned int res_bitshift; /* from the LEFT! */ unsigned int res_size; /* size of resource map in bytes */ -#if SBA_HINT_SUPPORT +#ifdef SBA_HINT_SUPPORT /* FIXME : DMA HINTs not used */ unsigned long hint_mask_pdir; /* bits used for DMA hints */ unsigned int hint_shift_pdir; @@ -294,7 +294,7 @@ static unsigned long piranha_bad_128k = 0; /* Looks nice and keeps the compiler happy */ #define SBA_DEV(d) ((struct sba_device *) (d)) -#if SBA_AGP_SUPPORT +#ifdef SBA_AGP_SUPPORT static int reserve_sba_gart = 1; #endif @@ -314,7 +314,7 @@ static int reserve_sba_gart = 1; #define WRITE_REG32(val, addr) __raw_writel(cpu_to_le32(val), addr) #define WRITE_REG64(val, addr) __raw_writeq(cpu_to_le64(val), addr) -#ifdef __LP64__ +#ifdef CONFIG_64BIT #define READ_REG(addr) READ_REG64(addr) #define WRITE_REG(value, addr) WRITE_REG64(value, addr) #else @@ -324,7 +324,7 @@ static int reserve_sba_gart = 1; #ifdef DEBUG_SBA_INIT -/* NOTE: When __LP64__ isn't defined, READ_REG64() is two 32-bit reads */ +/* NOTE: When CONFIG_64BIT isn't defined, READ_REG64() is two 32-bit reads */ /** * sba_dump_ranges - debugging only - print ranges assigned to this IOA @@ -364,7 +364,7 @@ static void sba_dump_tlb(void __iomem *hpa) #else #define sba_dump_ranges(x) #define sba_dump_tlb(x) -#endif +#endif /* DEBUG_SBA_INIT */ #ifdef ASSERT_PDIR_SANITY @@ -674,7 +674,7 @@ sba_free_range(struct ioc *ioc, dma_addr_t iova, size_t size) * ***************************************************************/ -#if SBA_HINT_SUPPORT +#ifdef SBA_HINT_SUPPORT #define SBA_DMA_HINT(ioc, val) ((val) << (ioc)->hint_shift_pdir) #endif @@ -743,9 +743,8 @@ sba_io_pdir_entry(u64 *pdir_ptr, space_t sid, unsigned long vba, * (bit #61, big endian), we have to flush and sync every time * IO-PDIR is changed in Ike/Astro. */ - if (ioc_needs_fdc) { - asm volatile("fdc 0(%%sr1,%0)\n\tsync" : : "r" (pdir_ptr)); - } + if (ioc_needs_fdc) + asm volatile("fdc %%r0(%0)" : : "r" (pdir_ptr)); } @@ -769,42 +768,57 @@ static SBA_INLINE void sba_mark_invalid(struct ioc *ioc, dma_addr_t iova, size_t byte_cnt) { u32 iovp = (u32) SBA_IOVP(ioc,iova); - - /* Even though this is a big-endian machine, the entries - ** in the iopdir are little endian. That's why we clear the byte - ** at +7 instead of at +0. - */ - int off = PDIR_INDEX(iovp)*sizeof(u64)+7; + u64 *pdir_ptr = &ioc->pdir_base[PDIR_INDEX(iovp)]; #ifdef ASSERT_PDIR_SANITY - /* Assert first pdir entry is set */ - if (0x80 != (((u8 *) ioc->pdir_base)[off])) { + /* Assert first pdir entry is set. + ** + ** Even though this is a big-endian machine, the entries + ** in the iopdir are little endian. That's why we look at + ** the byte at +7 instead of at +0. + */ + if (0x80 != (((u8 *) pdir_ptr)[7])) { sba_dump_pdir_entry(ioc,"sba_mark_invalid()", PDIR_INDEX(iovp)); } #endif - if (byte_cnt <= IOVP_SIZE) + if (byte_cnt > IOVP_SIZE) { - iovp |= IOVP_SHIFT; /* set "size" field for PCOM */ +#if 0 + unsigned long entries_per_cacheline = ioc_needs_fdc ? + L1_CACHE_ALIGN(((unsigned long) pdir_ptr)) + - (unsigned long) pdir_ptr; + : 262144; +#endif - /* - ** clear I/O PDIR entry "valid" bit - ** Do NOT clear the rest - save it for debugging. - ** We should only clear bits that have previously - ** been enabled. - */ - ((u8 *)(ioc->pdir_base))[off] = 0; - } else { - u32 t = get_order(byte_cnt) + PAGE_SHIFT; + /* set "size" field for PCOM */ + iovp |= get_order(byte_cnt) + PAGE_SHIFT; - iovp |= t; do { /* clear I/O Pdir entry "valid" bit first */ - ((u8 *)(ioc->pdir_base))[off] = 0; - off += sizeof(u64); + ((u8 *) pdir_ptr)[7] = 0; + if (ioc_needs_fdc) { + asm volatile("fdc %%r0(%0)" : : "r" (pdir_ptr)); +#if 0 + entries_per_cacheline = L1_CACHE_SHIFT - 3; +#endif + } + pdir_ptr++; byte_cnt -= IOVP_SIZE; - } while (byte_cnt > 0); - } + } while (byte_cnt > IOVP_SIZE); + } else + iovp |= IOVP_SHIFT; /* set "size" field for PCOM */ + + /* + ** clear I/O PDIR entry "valid" bit. + ** We have to R/M/W the cacheline regardless how much of the + ** pdir entry that we clobber. + ** The rest of the entry would be useful for debugging if we + ** could dump core on HPMC. + */ + ((u8 *) pdir_ptr)[7] = 0; + if (ioc_needs_fdc) + asm volatile("fdc %%r0(%0)" : : "r" (pdir_ptr)); WRITE_REG( SBA_IOVA(ioc, iovp, 0, 0), ioc->ioc_hpa+IOC_PCOM); } @@ -819,18 +833,29 @@ sba_mark_invalid(struct ioc *ioc, dma_addr_t iova, size_t byte_cnt) static int sba_dma_supported( struct device *dev, u64 mask) { struct ioc *ioc; + if (dev == NULL) { printk(KERN_ERR MODULE_NAME ": EISA/ISA/et al not supported\n"); BUG(); return(0); } - ioc = GET_IOC(dev); + /* Documentation/DMA-mapping.txt tells drivers to try 64-bit first, + * then fall back to 32-bit if that fails. + * We are just "encouraging" 32-bit DMA masks here since we can + * never allow IOMMU bypass unless we add special support for ZX1. + */ + if (mask > ~0U) + return 0; - /* check if mask is > than the largest IO Virt Address */ + ioc = GET_IOC(dev); - return((int) (mask >= (ioc->ibase + - (ioc->pdir_size / sizeof(u64) * IOVP_SIZE) ))); + /* + * check if mask is >= than the current max IO Virt Address + * The max IO Virt address will *always* < 30 bits. + */ + return((int)(mask >= (ioc->ibase - 1 + + (ioc->pdir_size / sizeof(u64) * IOVP_SIZE) ))); } @@ -898,11 +923,17 @@ sba_map_single(struct device *dev, void *addr, size_t size, size -= IOVP_SIZE; pdir_start++; } - /* form complete address */ + + /* force FDC ops in io_pdir_entry() to be visible to IOMMU */ + if (ioc_needs_fdc) + asm volatile("sync" : : ); + #ifdef ASSERT_PDIR_SANITY sba_check_pdir(ioc,"Check after sba_map_single()"); #endif spin_unlock_irqrestore(&ioc->res_lock, flags); + + /* form complete address */ return SBA_IOVA(ioc, iovp, offset, DEFAULT_DMA_HINT_REG); } @@ -958,12 +989,19 @@ sba_unmap_single(struct device *dev, dma_addr_t iova, size_t size, d--; } ioc->saved_cnt = 0; + READ_REG(ioc->ioc_hpa+IOC_PCOM); /* flush purges */ } #else /* DELAYED_RESOURCE_CNT == 0 */ sba_free_range(ioc, iova, size); + + /* If fdc's were issued, force fdc's to be visible now */ + if (ioc_needs_fdc) + asm volatile("sync" : : ); + READ_REG(ioc->ioc_hpa+IOC_PCOM); /* flush purges */ #endif /* DELAYED_RESOURCE_CNT == 0 */ + spin_unlock_irqrestore(&ioc->res_lock, flags); /* XXX REVISIT for 2.5 Linux - need syncdma for zero-copy support. @@ -1106,6 +1144,10 @@ sba_map_sg(struct device *dev, struct scatterlist *sglist, int nents, */ filled = iommu_fill_pdir(ioc, sglist, nents, 0, sba_io_pdir_entry); + /* force FDC ops in io_pdir_entry() to be visible to IOMMU */ + if (ioc_needs_fdc) + asm volatile("sync" : : ); + #ifdef ASSERT_PDIR_SANITY if (sba_check_pdir(ioc,"Check after sba_map_sg()")) { @@ -1234,8 +1276,10 @@ sba_alloc_pdir(unsigned int pdir_size) unsigned long pdir_order = get_order(pdir_size); pdir_base = __get_free_pages(GFP_KERNEL, pdir_order); - if (NULL == (void *) pdir_base) - panic("sba_ioc_init() could not allocate I/O Page Table\n"); + if (NULL == (void *) pdir_base) { + panic("%s() could not allocate I/O Page Table\n", + __FUNCTION__); + } /* If this is not PA8700 (PCX-W2) ** OR newer than ver 2.2 @@ -1353,7 +1397,7 @@ sba_ioc_init_pluto(struct parisc_device *sba, struct ioc *ioc, int ioc_num) u32 iova_space_mask; u32 iova_space_size; int iov_order, tcnfg; -#if SBA_AGP_SUPPORT +#ifdef SBA_AGP_SUPPORT int agp_found = 0; #endif /* @@ -1390,7 +1434,7 @@ sba_ioc_init_pluto(struct parisc_device *sba, struct ioc *ioc, int ioc_num) DBG_INIT("%s() pdir %p size %x\n", __FUNCTION__, ioc->pdir_base, ioc->pdir_size); -#if SBA_HINT_SUPPORT +#ifdef SBA_HINT_SUPPORT ioc->hint_shift_pdir = iov_order + PAGE_SHIFT; ioc->hint_mask_pdir = ~(0x3 << (iov_order + PAGE_SHIFT)); @@ -1414,7 +1458,7 @@ sba_ioc_init_pluto(struct parisc_device *sba, struct ioc *ioc, int ioc_num) WRITE_REG(ioc->imask, ioc->ioc_hpa + IOC_IMASK); -#ifdef __LP64__ +#ifdef CONFIG_64BIT /* ** Setting the upper bits makes checking for bypass addresses ** a little faster later on. @@ -1447,7 +1491,7 @@ sba_ioc_init_pluto(struct parisc_device *sba, struct ioc *ioc, int ioc_num) */ WRITE_REG(ioc->ibase | 31, ioc->ioc_hpa + IOC_PCOM); -#if SBA_AGP_SUPPORT +#ifdef SBA_AGP_SUPPORT /* ** If an AGP device is present, only use half of the IOV space ** for PCI DMA. Unfortunately we can't know ahead of time @@ -1499,11 +1543,9 @@ sba_ioc_init(struct parisc_device *sba, struct ioc *ioc, int ioc_num) if (iova_space_size < (1 << (20 - PAGE_SHIFT))) { iova_space_size = 1 << (20 - PAGE_SHIFT); } -#ifdef __LP64__ else if (iova_space_size > (1 << (30 - PAGE_SHIFT))) { iova_space_size = 1 << (30 - PAGE_SHIFT); } -#endif /* ** iova space must be log2() in size. @@ -1529,7 +1571,7 @@ sba_ioc_init(struct parisc_device *sba, struct ioc *ioc, int ioc_num) DBG_INIT("%s() pdir %p size %x\n", __FUNCTION__, ioc->pdir_base, pdir_size); -#if SBA_HINT_SUPPORT +#ifdef SBA_HINT_SUPPORT /* FIXME : DMA HINTs not used */ ioc->hint_shift_pdir = iov_order + PAGE_SHIFT; ioc->hint_mask_pdir = ~(0x3 << (iov_order + PAGE_SHIFT)); -- cgit v1.2.3-59-g8ed1b