/* * CDDL HEADER START * * The contents of this file are subject to the terms of the * Common Development and Distribution License (the "License"). * You may not use this file except in compliance with the License. * * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE * or http://www.opensolaris.org/os/licensing. * See the License for the specific language governing permissions * and limitations under the License. * * When distributing Covered Code, include this CDDL HEADER in each * file and include the License file at usr/src/OPENSOLARIS.LICENSE. * If applicable, add the following below this CDDL HEADER, with the * fields enclosed by brackets "[]" replaced with your own identifying * information: Portions Copyright [yyyy] [name of copyright owner] * * CDDL HEADER END */ /* * Copyright (c) 1990, 2010, Oracle and/or its affiliates. All rights reserved. */ /* * Copyright 2014 Garrett D'Amore */ #ifndef _SYS_DDIDMAREQ_H #define _SYS_DDIDMAREQ_H #ifdef __cplusplus extern "C" { #endif /* * Memory Objects * * Definitions of structures that can describe * an object that can be mapped for DMA. */ /* * Structure describing a virtual address */ struct v_address { caddr_t v_addr; /* base virtual address */ struct as *v_as; /* pointer to address space */ void *v_priv; /* priv data for shadow I/O */ }; /* * Structure describing a page-based address */ struct pp_address { /* * A pointer to a circularly linked list of page structures. */ struct page *pp_pp; uint_t pp_offset; /* offset within first page */ }; /* * Structure to describe a physical memory address. */ struct phy_address { ulong_t p_addr; /* base physical address */ ulong_t p_memtype; /* memory type */ }; /* * Structure to describe an array DVMA addresses. * Under normal circumstances, dv_nseg will be 1. * dvs_start is always page aligned. */ struct dvma_address { size_t dv_off; size_t dv_nseg; struct dvmaseg { uint64_t dvs_start; size_t dvs_len; } *dv_seg; }; /* * A union of all of the above structures. * * This union describes the relationship between * the kind of an address description and an object. */ typedef union { struct v_address virt_obj; /* Some virtual address */ struct pp_address pp_obj; /* Some page-based address */ struct phy_address phys_obj; /* Some physical address */ struct dvma_address dvma_obj; } ddi_dma_aobj_t; /* * DMA object types - used to select how the object * being mapped is being addressed by the IU. */ typedef enum { DMA_OTYP_VADDR = 0, /* enforce starting value of zero */ DMA_OTYP_PAGES, DMA_OTYP_PADDR, DMA_OTYP_BUFVADDR, DMA_OTYP_DVADDR } ddi_dma_atyp_t; /* * A compact package to describe an object that is to be mapped for DMA. */ typedef struct { uint_t dmao_size; /* size, in bytes, of the object */ ddi_dma_atyp_t dmao_type; /* type of object */ ddi_dma_aobj_t dmao_obj; /* the object described */ } ddi_dma_obj_t; /* * DMA addressing limits. * * This structure describes the constraints that a particular device's * DMA engine has to its parent so that the parent may correctly set * things up for a DMA mapping. Each parent may in turn modify the * constraints listed in a DMA request structure in order to describe * to its parent any changed or additional constraints. The rules * are that each parent may modify a constraint in order to further * constrain things (e.g., picking a more limited address range than * that permitted by the child), but that the parent may not ignore * a child's constraints. * * A particular constraint that we do *not* address is whether or not * a requested mapping is too large for a DMA engine's counter to * correctly track. It is still up to each driver to explicitly handle * transfers that are too large for its own hardware to deal with directly. * * The mapping routines that are cognizant of this structure will * copy any user defined limits structure if they need to modify * the fields (as alluded to above). * * A note as to how to define constraints: * * How you define the constraints for your device depends on how you * define your device. For example, you may have an SBus card with a * device on it that address only the bottom 16mb of virtual DMA space. * However, if the card also has ancillary circuitry that pulls the high 8 * bits of address lines high, the more correct expression for your device * is that it address [0xff000000..0xffffffff] rather than [0..0x00ffffff]. */ #if defined(__sparc) typedef struct ddi_dma_lim { /* * Low range of 32 bit addressing capability. */ uint_t dlim_addr_lo; /* * Upper inclusive bound of addressing capability. It is an * inclusive boundary limit to allow for the addressing range * [0..0xffffffff] to be specified in preference to [0..0]. */ uint_t dlim_addr_hi; /* * Inclusive upper bound with which The DMA engine's counter acts as * a register. * * This handles the case where an upper portion of a DMA address * register is a latch instead of being a full 32 bit register * (e.g., the upper 8 bits may remain constant while the lower * 24 bits are the real address register). * * This essentially gives a hint about segment limitations * to the mapping routines. */ uint_t dlim_cntr_max; /* * DMA burst sizes. * * At the time of a mapping request, this tag defines the possible * DMA burst cycle sizes that the requestor's DMA engine can * emit. The format of the data is binary encoding of burst sizes * assumed to be powers of two. That is, if a DMA engine is capable * of doing 1, 2, 4 and 16 byte transfers, the encoding would be 0x17. * * As the mapping request is handled by intervening nexi, the * burstsizes value may be modified. Prior to enabling DMA for * the specific device, the driver that owns the DMA engine should * check (via ddi_dma_burstsizes(9F)) what the allowed burstsizes * have become and program their DMA engine appropriately. */ uint_t dlim_burstsizes; /* * Minimum effective DMA transfer size, in units of bytes. * * This value specifies the minimum effective granularity of the * DMA engine. It is distinct from dlim_burtsizes in that it * describes the minimum amount of access a DMA transfer will * effect. dlim_burtsizes describes in what electrical fashion * the DMA engine might perform its accesses, while dlim_minxfer * describes the minimum amount of memory that can be touched by * the DMA transfer. * * As the mapping request is handled by intervening nexi, the * dlim_minxfer value may be modifed contingent upon the presence * (and use) of I/O caches and DMA write buffers in between the * DMA engine and the object that DMA is being performed on. * */ uint_t dlim_minxfer; /* * Expected average data rate for this DMA engine * while transferring data. * * This is used as a hint for a number of operations that might * want to know the possible optimal latency requirements of this * device. A value of zero will be interpreted as a 'do not care'. */ uint_t dlim_dmaspeed; } ddi_dma_lim_t; #elif defined(__x86) /* * values for dlim_minxfer */ #define DMA_UNIT_8 1 #define DMA_UNIT_16 2 #define DMA_UNIT_32 4 /* * Version number */ #define DMALIM_VER0 ((0x86000000) + 0) typedef struct ddi_dma_lim { /* * Low range of 32 bit addressing capability. */ uint_t dlim_addr_lo; /* * Upper Inclusive bound of 32 bit addressing capability. * * The ISA nexus restricts this to 0x00ffffff, since this bus has * only 24 address lines. This enforces the 16 Mb address limitation. * The EISA nexus restricts this to 0xffffffff. */ uint_t dlim_addr_hi; /* * DMA engine counter not used; set to 0 */ uint_t dlim_cntr_max; /* * DMA burst sizes not used; set to 1 */ uint_t dlim_burstsizes; /* * Minimum effective DMA transfer size. * * This value specifies the minimum effective granularity of the * DMA engine. It is distinct from dlim_burstsizes in that it * describes the minimum amount of access a DMA transfer will * effect. dlim_burstsizes describes in what electrical fashion * the DMA engine might perform its accesses, while dlim_minxfer * describes the minimum amount of memory that can be touched by * the DMA transfer. * * This value also implies the required address alignment. * The number of bytes transferred is assumed to be * dlim_minxfer * (DMA engine count) * * It should be set to DMA_UNIT_8, DMA_UNIT_16, or DMA_UNIT_32. */ uint_t dlim_minxfer; /* * Expected average data rate for this DMA engine * while transferring data. * * This is used as a hint for a number of operations that might * want to know the possible optimal latency requirements of this * device. A value of zero will be interpreted as a 'do not care'. */ uint_t dlim_dmaspeed; /* * Version number of this structure */ uint_t dlim_version; /* = 0x86 << 24 + 0 */ /* * Inclusive upper bound with which the DMA engine's Address acts as * a register. * This handles the case where an upper portion of a DMA address * register is a latch instead of being a full 32 bit register * (e.g., the upper 16 bits remain constant while the lower 16 bits * are incremented for each DMA transfer). * * The ISA nexus restricts only 3rd-party DMA requests to 0x0000ffff, * since the ISA DMA engine has a 16-bit register for low address and * an 8-bit latch for high address. This enforces the first 64 Kb * limitation (address boundary). * The EISA nexus restricts only 3rd-party DMA requests to 0xffffffff. */ uint_t dlim_adreg_max; /* * Maximum transfer count that the DMA engine can handle. * * The ISA nexus restricts only 3rd-party DMA requests to 0x0000ffff, * since the ISA DMA engine has a 16-bit register for counting. * This enforces the other 64 Kb limitation (count size). * The EISA nexus restricts only 3rd-party DMA requests to 0x00ffffff, * since the EISA DMA engine has a 24-bit register for counting. * * This transfer count limitation is a per segment limitation. * It can also be used to restrict the size of segments. * * This is used as a bit mask, so it must be a power of 2, minus 1. */ uint_t dlim_ctreg_max; /* * Granularity of DMA transfer, in units of bytes. * * Breakup sizes must be multiples of this value. * If no scatter/gather capabilty is specified, then the size of * each DMA transfer must be a multiple of this value. * * If there is scatter/gather capability, then a single cookie cannot * be smaller in size than the minimum xfer value, and may be less * than the granularity value. The total transfer length of the * scatter/gather list should be a multiple of the granularity value; * use dlim_sgllen to specify the length of the scatter/gather list. * * This value should be equal to the sector size of the device. */ uint_t dlim_granular; /* * Length of scatter/gather list * * This value specifies the number of segments or cookies that a DMA * engine can consume in one i/o request to the device. For 3rd-party * DMA that uses the bus nexus this should be set to 1. Devices with * 1st-party DMA capability should specify the number of entries in * its scatter/gather list. The breakup routine will ensure that each * group of dlim_sgllen cookies (within a DMA window) will have a * total transfer length that is a multiple of dlim_granular. * * < 0 : tbd * = 0 : breakup is for PIO. * = 1 : breakup is for DMA engine with no scatter/gather * capability. * >= 2 : breakup is for DMA engine with scatter/gather * capability; value is max number of entries in list. * * Note that this list length is not dependent on the DMA window * size. The size of the DMA window is based on resources consumed, * such as intermediate buffers. Several s/g lists may exist within * a window. But the end of a window does imply the end of the s/g * list. */ short dlim_sgllen; /* * Size of device i/o request * * This value indicates the maximum number of bytes the device * can transmit/receive for one i/o command. This limitation is * significant ony if it is less than (dlim_ctreg_max * dlim_sgllen). */ uint_t dlim_reqsize; } ddi_dma_lim_t; #else #error "struct ddi_dma_lim not defined for this architecture" #endif /* defined(__sparc) */ /* * Flags definition for dma_attr_flags */ /* * return physical DMA address on platforms * which support DVMA */ #define DDI_DMA_FORCE_PHYSICAL 0x0100 /* * An error will be flagged for DMA data path errors */ #define DDI_DMA_FLAGERR 0x200 /* * Enable relaxed ordering */ #define DDI_DMA_RELAXED_ORDERING 0x400 /* * Consolidation private x86 only flag which will cause a bounce buffer * (paddr < dma_attr_seg) to be used if the buffer passed to the bind * operation contains pages both above and below dma_attr_seg. If this flag * is set, dma_attr_seg must be <= dma_attr_addr_hi. */ #define _DDI_DMA_BOUNCE_ON_SEG 0x8000 #define DMA_ATTR_V0 0 #define DMA_ATTR_VERSION DMA_ATTR_V0 typedef struct ddi_dma_attr { uint_t dma_attr_version; /* version number */ uint64_t dma_attr_addr_lo; /* low DMA address range */ uint64_t dma_attr_addr_hi; /* high DMA address range */ uint64_t dma_attr_count_max; /* DMA counter register */ uint64_t dma_attr_align; /* DMA address alignment */ uint_t dma_attr_burstsizes; /* DMA burstsizes */ uint32_t dma_attr_minxfer; /* min effective DMA size */ uint64_t dma_attr_maxxfer; /* max DMA xfer size */ uint64_t dma_attr_seg; /* segment boundary */ int dma_attr_sgllen; /* s/g length */ uint32_t dma_attr_granular; /* granularity of device */ uint_t dma_attr_flags; /* Bus specific DMA flags */ } ddi_dma_attr_t; /* * Handy macro to set a maximum bit value (should be elsewhere) * * Clear off all bits lower then 'mybit' in val; if there are no * bits higher than or equal to mybit in val then set mybit. Assumes * mybit equals some power of 2 and is not zero. */ #define maxbit(val, mybit) \ ((val) & ~((mybit)-1)) | ((((val) & ~((mybit)-1)) == 0) ? (mybit) : 0) /* * Handy macro to set a minimum bit value (should be elsewhere) * * Clear off all bits higher then 'mybit' in val; if there are no * bits lower than or equal to mybit in val then set mybit. Assumes * mybit equals some pow2 and is not zero. */ #define minbit(val, mybit) \ (((val)&((mybit)|((mybit)-1))) | \ ((((val) & ((mybit)-1)) == 0) ? (mybit) : 0)) /* * Structure of a request to map an object for DMA. */ typedef struct ddi_dma_req { /* * Caller's DMA engine constraints. * * If there are no particular constraints to the caller's DMA * engine, this field may be set to NULL. The implementation DMA * setup functions will then select a set of standard beginning * constraints. * * In either case, as the mapping proceeds, the initial DMA * constraints may become more restrictive as each intervening * nexus might add further restrictions. */ ddi_dma_lim_t *dmar_limits; /* * Contains the information passed to the DMA mapping allocation * routine(s). */ uint_t dmar_flags; /* * Callback function. A caller of the DMA mapping functions must * specify by filling in this field whether the allocation routines * can sleep awaiting mapping resources, must *not* sleep awaiting * resources, or may *not* sleep awaiting any resources and must * call the function specified by dmar_fp with the the argument * dmar_arg when resources might have become available at a future * time. */ int (*dmar_fp)(); caddr_t dmar_arg; /* Callback function argument */ /* * Description of the object to be mapped for DMA. * Must be last in this structure in case that the * union ddi_dma_obj_t changes in the future. */ ddi_dma_obj_t dmar_object; } ddi_dma_req_t; /* * Defines for the DMA mapping allocation functions * * If a DMA callback funtion is set to anything other than the following * defines then it is assumed that one wishes a callback and is providing * a function address. */ #ifdef __STDC__ #define DDI_DMA_DONTWAIT ((int (*)(caddr_t))0) #define DDI_DMA_SLEEP ((int (*)(caddr_t))1) #else #define DDI_DMA_DONTWAIT ((int (*)())0) #define DDI_DMA_SLEEP ((int (*)())1) #endif /* * Return values from callback functions. */ #define DDI_DMA_CALLBACK_RUNOUT 0 #define DDI_DMA_CALLBACK_DONE 1 /* * Flag definitions for the allocation functions. */ #define DDI_DMA_WRITE 0x0001 /* Direction memory --> IO */ #define DDI_DMA_READ 0x0002 /* Direction IO --> memory */ #define DDI_DMA_RDWR (DDI_DMA_READ | DDI_DMA_WRITE) /* * If possible, establish a MMU redzone after the mapping (to protect * against cheap DMA hardware that might get out of control). */ #define DDI_DMA_REDZONE 0x0004 /* * A partial allocation is allowed. That is, if the size of the object * exceeds the mapping resources available, only map a portion of the * object and return status indicating that this took place. The caller * can use the functions ddi_dma_numwin(9F) and ddi_dma_getwin(9F) to * change, at a later point, the actual mapped portion of the object. * * The mapped portion begins at offset 0 of the object. * */ #define DDI_DMA_PARTIAL 0x0008 /* * Map the object for byte consistent access. Note that explicit * synchronization (via ddi_dma_sync(9F)) will still be required. * Consider this flag to be a hint to the mapping routines as to * the intended use of the mapping. * * Normal data transfers can be usually consider to use 'streaming' * modes of operations. They start at a specific point, transfer a * fairly large amount of data sequentially, and then stop (usually * on a well aligned boundary). * * Control mode data transfers (for memory resident device control blocks, * e.g., ethernet message descriptors) do not access memory in such * a streaming sequential fashion. Instead, they tend to modify a few * words or bytes, move around and maybe modify a few more. * * There are many machine implementations that make this difficult to * control in a generic and seamless fashion. Therefore, explicit synch- * ronization steps (via ddi_dma_sync(9F)) are still required (even if you * ask for a byte-consistent mapping) in order to make the view of the * memory object shared between a CPU and a DMA master in consistent. * However, judicious use of this flag can give sufficient hints to * the mapping routines to attempt to pick the most efficacious mapping * such that the synchronization steps are as efficient as possible. * */ #define DDI_DMA_CONSISTENT 0x0010 /* * Some DMA mappings have to be 'exclusive' access. */ #define DDI_DMA_EXCLUSIVE 0x0020 /* * Sequential, unidirectional, block-sized and block aligned transfers */ #define DDI_DMA_STREAMING 0x0040 /* * Support for 64-bit SBus devices */ #define DDI_DMA_SBUS_64BIT 0x2000 /* * Return values from the mapping allocation functions. */ /* * succeeded in satisfying request */ #define DDI_DMA_MAPPED 0 /* * Mapping is legitimate (for advisory calls). */ #define DDI_DMA_MAPOK 0 /* * Succeeded in mapping a portion of the request. */ #define DDI_DMA_PARTIAL_MAP 1 /* * indicates end of window/segment list */ #define DDI_DMA_DONE 2 /* * No resources to map request. */ #define DDI_DMA_NORESOURCES -1 /* * Can't establish a mapping to the specified object * (no specific reason). */ #define DDI_DMA_NOMAPPING -2 /* * The request is too big to be mapped. */ #define DDI_DMA_TOOBIG -3 /* * The request is too small to be mapped. */ #define DDI_DMA_TOOSMALL -4 /* * The request cannot be mapped because the object * is locked against mapping by another DMA master. */ #define DDI_DMA_LOCKED -5 /* * The request cannot be mapped because the limits * structure has bogus values. */ #define DDI_DMA_BADLIMITS -6 /* * the segment/window pointer is stale */ #define DDI_DMA_STALE -7 /* * The system can't allocate DMA resources using * the given DMA attributes */ #define DDI_DMA_BADATTR -8 /* * A DMA handle is already used for a DMA */ #define DDI_DMA_INUSE -9 /* * DVMA disabled or not supported. use physical DMA */ #define DDI_DMA_USE_PHYSICAL -10 /* * In order for the access to a memory object to be consistent * between a device and a CPU, the function ddi_dma_sync(9F) * must be called upon the DMA handle. The following flags * define whose view of the object should be made consistent. * There are different flags here because on different machines * there are definite performance implications of how long * such synchronization takes. * * DDI_DMA_SYNC_FORDEV makes all device references to the object * mapped by the DMA handle up to date. It should be used by a * driver after a cpu modifies the memory object (over the range * specified by the other arguments to the ddi_dma_sync(9F) call). * * DDI_DMA_SYNC_FORCPU makes all cpu references to the object * mapped by the DMA handle up to date. It should be used * by a driver after the receipt of data from the device to * the memory object is done (over the range specified by * the other arguments to the ddi_dma_sync(9F) call). * * If the only mapping that concerns the driver is one for the * kernel (such as memory allocated by ddi_iopb_alloc(9F)), the * flag DDI_DMA_SYNC_FORKERNEL can be used. This is a hint to the * system that if it can synchronize the kernel's view faster * that the CPU's view, it can do so, otherwise it acts the * same as DDI_DMA_SYNC_FORCPU. DDI_DMA_SYNC_FORKERNEL might * speed up the synchronization of kernel mappings in case of * non IO-coherent CPU caches. */ #define DDI_DMA_SYNC_FORDEV 0x0 #define DDI_DMA_SYNC_FORCPU 0x1 #define DDI_DMA_SYNC_FORKERNEL 0x2 /* * Bus nexus control functions for DMA */ /* * Control operations, defined here so that devops.h can be included * by drivers without having to include a specific SYSDDI implementation * header file. */ enum ddi_dma_ctlops { DDI_DMA_FREE, /* obsolete - do not use */ DDI_DMA_SYNC, /* obsolete - do not use */ DDI_DMA_HTOC, /* obsolete - do not use */ DDI_DMA_KVADDR, /* obsolete - do not use */ DDI_DMA_MOVWIN, /* obsolete - do not use */ DDI_DMA_REPWIN, /* obsolete - do not use */ DDI_DMA_GETERR, /* obsolete - do not use */ DDI_DMA_COFF, /* obsolete - do not use */ DDI_DMA_NEXTWIN, /* obsolete - do not use */ DDI_DMA_NEXTSEG, /* obsolete - do not use */ DDI_DMA_SEGTOC, /* obsolete - do not use */ DDI_DMA_RESERVE, /* reserve some DVMA range */ DDI_DMA_RELEASE, /* free preallocated DVMA range */ DDI_DMA_RESETH, /* obsolete - do not use */ DDI_DMA_CKSYNC, /* obsolete - do not use */ DDI_DMA_IOPB_ALLOC, /* obsolete - do not use */ DDI_DMA_IOPB_FREE, /* obsolete - do not use */ DDI_DMA_SMEM_ALLOC, /* obsolete - do not use */ DDI_DMA_SMEM_FREE, /* obsolete - do not use */ DDI_DMA_SET_SBUS64, /* 64 bit SBus support */ DDI_DMA_REMAP, /* remap DVMA buffers after relocation */ /* * control ops for DMA engine on motherboard */ DDI_DMA_E_ACQUIRE, /* get channel for exclusive use */ DDI_DMA_E_FREE, /* release channel */ DDI_DMA_E_1STPTY, /* setup channel for 1st party DMA */ DDI_DMA_E_GETCB, /* get control block for DMA engine */ DDI_DMA_E_FREECB, /* free control blk for DMA engine */ DDI_DMA_E_PROG, /* program channel of DMA engine */ DDI_DMA_E_SWSETUP, /* setup channel for software control */ DDI_DMA_E_SWSTART, /* software operation of DMA channel */ DDI_DMA_E_ENABLE, /* enable channel of DMA engine */ DDI_DMA_E_STOP, /* stop a channel of DMA engine */ DDI_DMA_E_DISABLE, /* disable channel of DMA engine */ DDI_DMA_E_GETCNT, /* get remaining xfer count */ DDI_DMA_E_GETLIM, /* obsolete - do not use */ DDI_DMA_E_GETATTR /* get DMA engine attributes */ }; /* * Cache attribute flags: * * IOMEM_DATA_CACHED * The CPU can cache the data it fetches and push it to memory at a later * time. This is the default attribute and used if no cache attributes is * specified. * * IOMEM_DATA_UC_WR_COMBINE * The CPU never caches the data but writes may occur out of order or be * combined. It implies re-ordering. * * IOMEM_DATA_UNCACHED * The CPU never caches the data and has uncacheable access to memory. * It also implies strict ordering. * * The cache attributes are mutually exclusive, and any combination of the * values leads to a failure. On the sparc architecture, only IOMEM_DATA_CACHED * is meaningful, but others lead to a failure. */ #define IOMEM_DATA_CACHED 0x10000 /* data is cached */ #define IOMEM_DATA_UC_WR_COMBINE 0x20000 /* data is not cached, but */ /* writes might be combined */ #define IOMEM_DATA_UNCACHED 0x40000 /* data is not cached. */ #define IOMEM_DATA_MASK 0xF0000 /* cache attrs mask */ /* * Check if either uncacheable or write-combining specified. (those flags are * mutually exclusive) This macro is used to override hat attributes if either * one is set. */ #define OVERRIDE_CACHE_ATTR(attr) \ (attr & (IOMEM_DATA_UNCACHED | IOMEM_DATA_UC_WR_COMBINE)) /* * Get the cache attribute from flags. If there is no attributes, * return IOMEM_DATA_CACHED (default attribute). */ #define IOMEM_CACHE_ATTR(flags) \ ((flags & IOMEM_DATA_MASK) ? (flags & IOMEM_DATA_MASK) : \ IOMEM_DATA_CACHED) #ifdef __cplusplus } #endif #endif /* _SYS_DDIDMAREQ_H */