Home | History | Annotate | Line # | Download | only in dev
mm.c revision 1.25 
      1 /*	$NetBSD: mm.c,v 1.25 2024/08/25 11:29:38 uwe Exp $	*/
      2 
      3 /*-
      4  * Copyright (c) 2002, 2008, 2010 The NetBSD Foundation, Inc.
      5  * All rights reserved.
      6  *
      7  * This code is derived from software contributed to The NetBSD Foundation
      8  * by Christos Zoulas, Joerg Sonnenberger and Mindaugas Rasiukevicius.
      9  *
     10  * Redistribution and use in source and binary forms, with or without
     11  * modification, are permitted provided that the following conditions
     12  * are met:
     13  * 1. Redistributions of source code must retain the above copyright
     14  *    notice, this list of conditions and the following disclaimer.
     15  * 2. Redistributions in binary form must reproduce the above copyright
     16  *    notice, this list of conditions and the following disclaimer in the
     17  *    documentation and/or other materials provided with the distribution.
     18  *
     19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
     20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
     21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
     22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
     23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
     24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
     25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
     26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
     27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
     28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
     29  * POSSIBILITY OF SUCH DAMAGE.
     30  */
     31 
     32 /*
     33  * Special /dev/{mem,kmem,zero,null} memory devices.
     34  */
     35 
     36 #include <sys/cdefs.h>
     37 __KERNEL_RCSID(0, "$NetBSD: mm.c,v 1.25 2024/08/25 11:29:38 uwe Exp $");
     38 
     39 #include "opt_compat_netbsd.h"
     40 
     41 #include <sys/param.h>
     42 #include <sys/conf.h>
     43 #include <sys/ioctl.h>
     44 #include <sys/mman.h>
     45 #include <sys/uio.h>
     46 #include <sys/termios.h>
     47 
     48 #include <dev/mm.h>
     49 
     50 #include <uvm/uvm_extern.h>
     51 
     52 static void *		dev_zero_page	__read_mostly;
     53 static kmutex_t		dev_mem_lock	__cacheline_aligned;
     54 static vaddr_t		dev_mem_addr	__read_mostly;
     55 
     56 static dev_type_open(mm_open);
     57 static dev_type_read(mm_readwrite);
     58 static dev_type_mmap(mm_mmap);
     59 static dev_type_ioctl(mm_ioctl);
     60 
     61 const struct cdevsw mem_cdevsw = {
     62 	.d_open = mm_open,
     63 	.d_close = nullclose,
     64 	.d_read = mm_readwrite,
     65 	.d_write = mm_readwrite,
     66 	.d_ioctl = mm_ioctl,
     67 	.d_stop = nostop,
     68 	.d_tty = notty,
     69 	.d_poll = nopoll,
     70 	.d_mmap = mm_mmap,
     71 	.d_kqfilter = nokqfilter,
     72 	.d_discard = nodiscard,
     73 	.d_flag = D_MPSAFE
     74 };
     75 
     76 #ifdef pmax	/* XXX */
     77 const struct cdevsw mem_ultrix_cdevsw = {
     78 	.d_open = nullopen,
     79 	.d_close = nullclose,
     80 	.d_read = mm_readwrite,
     81 	.d_write = mm_readwrite,
     82 	.d_ioctl = mm_ioctl,
     83 	.d_stop = nostop,
     84 	.d_tty = notty,
     85 	.d_poll = nopoll,
     86 	.d_mmap = mm_mmap,
     87 	.d_kqfilter = nokqfilter,
     88 	.d_discard = nodiscard,
     89 	.d_flag = D_MPSAFE
     90 };
     91 #endif
     92 
     93 static int
     94 mm_open(dev_t dev, int flag, int mode, struct lwp *l)
     95 {
     96 #ifdef __HAVE_MM_MD_OPEN
     97 	int error;
     98 	if ((error = mm_md_open(dev, flag, mode, l)) != 0)
     99 		return error;
    100 #endif
    101 	l->l_proc->p_flag |= PK_KMEM;
    102 	return 0;
    103 }
    104 
    105 /*
    106  * mm_init: initialize memory device driver.
    107  */
    108 void
    109 mm_init(void)
    110 {
    111 	vaddr_t pg;
    112 
    113 	mutex_init(&dev_mem_lock, MUTEX_DEFAULT, IPL_NONE);
    114 
    115 	/* Read-only zero-page. */
    116 	pg = uvm_km_alloc(kernel_map, PAGE_SIZE, 0, UVM_KMF_WIRED|UVM_KMF_ZERO);
    117 	KASSERT(pg != 0);
    118 	pmap_protect(pmap_kernel(), pg, pg + PAGE_SIZE, VM_PROT_READ);
    119 	pmap_update(pmap_kernel());
    120 	dev_zero_page = (void *)pg;
    121 
    122 #ifndef __HAVE_MM_MD_CACHE_ALIASING
    123 	/* KVA for mappings during I/O. */
    124 	dev_mem_addr = uvm_km_alloc(kernel_map, PAGE_SIZE, 0,
    125 	    UVM_KMF_VAONLY|UVM_KMF_WAITVA);
    126 	KASSERT(dev_mem_addr != 0);
    127 #else
    128 	dev_mem_addr = 0;
    129 #endif
    130 }
    131 
    132 
    133 /*
    134  * dev_mem_getva: get a special virtual address.  If architecture requires,
    135  * allocate VA according to PA, which avoids cache-aliasing issues.  Use a
    136  * constant, general mapping address otherwise.
    137  */
    138 static inline vaddr_t
    139 dev_mem_getva(paddr_t pa, int color)
    140 {
    141 #ifdef __HAVE_MM_MD_CACHE_ALIASING
    142 	return uvm_km_alloc(kernel_map, PAGE_SIZE,
    143 	    color & uvmexp.colormask,
    144 	    UVM_KMF_VAONLY | UVM_KMF_WAITVA | UVM_KMF_COLORMATCH);
    145 #else
    146 	return dev_mem_addr;
    147 #endif
    148 }
    149 
    150 static inline void
    151 dev_mem_relva(paddr_t pa, vaddr_t va)
    152 {
    153 #ifdef __HAVE_MM_MD_CACHE_ALIASING
    154 	uvm_km_free(kernel_map, va, PAGE_SIZE, UVM_KMF_VAONLY);
    155 #else
    156 	KASSERT(dev_mem_addr == va);
    157 #endif
    158 }
    159 
    160 /*
    161  * dev_kmem_readwrite: helper for DEV_MEM (/dev/mem) case of R/W.
    162  */
    163 static int
    164 dev_mem_readwrite(struct uio *uio, struct iovec *iov)
    165 {
    166 	paddr_t paddr;
    167 	vaddr_t vaddr;
    168 	vm_prot_t prot;
    169 	size_t len, offset;
    170 	bool have_direct;
    171 	int error;
    172 	int color = 0;
    173 
    174 	/* Check for wrap around. */
    175 	if ((uintptr_t)uio->uio_offset != uio->uio_offset) {
    176 		return EFAULT;
    177 	}
    178 	paddr = uio->uio_offset & ~PAGE_MASK;
    179 	prot = (uio->uio_rw == UIO_WRITE) ? VM_PROT_WRITE : VM_PROT_READ;
    180 	error = mm_md_physacc(paddr, prot);
    181 	if (error) {
    182 		return error;
    183 	}
    184 	offset = uio->uio_offset & PAGE_MASK;
    185 	len = MIN(uio->uio_resid, PAGE_SIZE - offset);
    186 
    187 #ifdef __HAVE_MM_MD_CACHE_ALIASING
    188 	have_direct = mm_md_page_color(paddr, &color);
    189 #else
    190 	have_direct = true;
    191 	color = 0;
    192 #endif
    193 
    194 #ifdef __HAVE_MM_MD_DIRECT_MAPPED_PHYS
    195 	/* Is physical address directly mapped?  Return VA. */
    196 	if (have_direct)
    197 		have_direct = mm_md_direct_mapped_phys(paddr, &vaddr);
    198 #else
    199 	vaddr = 0;
    200 	have_direct = false;
    201 #endif
    202 	if (!have_direct) {
    203 		/* Get a special virtual address. */
    204 		const vaddr_t va = dev_mem_getva(paddr, color);
    205 
    206 		/* Map selected KVA to physical address. */
    207 		mutex_enter(&dev_mem_lock);
    208 		pmap_kenter_pa(va, paddr, prot, 0);
    209 		pmap_update(pmap_kernel());
    210 
    211 		/* Perform I/O. */
    212 		vaddr = va + offset;
    213 		error = uiomove((void *)vaddr, len, uio);
    214 
    215 		/* Unmap, flush before unlock. */
    216 		pmap_kremove(va, PAGE_SIZE);
    217 		pmap_update(pmap_kernel());
    218 		mutex_exit(&dev_mem_lock);
    219 
    220 		/* "Release" the virtual address. */
    221 		dev_mem_relva(paddr, va);
    222 	} else {
    223 		/* Direct map, just perform I/O. */
    224 		vaddr += offset;
    225 		error = uiomove((void *)vaddr, len, uio);
    226 	}
    227 	return error;
    228 }
    229 
    230 /*
    231  * dev_kmem_readwrite: helper for DEV_KMEM (/dev/kmem) case of R/W.
    232  */
    233 static int
    234 dev_kmem_readwrite(struct uio *uio, struct iovec *iov)
    235 {
    236 	void *addr;
    237 	size_t len, offset;
    238 	vm_prot_t prot;
    239 	int error;
    240 	bool md_kva;
    241 
    242 	/* Check for wrap around. */
    243 	addr = (void *)(intptr_t)uio->uio_offset;
    244 	if ((uintptr_t)addr != uio->uio_offset) {
    245 		return EFAULT;
    246 	}
    247 	/*
    248 	 * Handle non-page aligned offset.
    249 	 * Otherwise, we operate in page-by-page basis.
    250 	 */
    251 	offset = uio->uio_offset & PAGE_MASK;
    252 	len = MIN(uio->uio_resid, PAGE_SIZE - offset);
    253 	prot = (uio->uio_rw == UIO_WRITE) ? VM_PROT_WRITE : VM_PROT_READ;
    254 
    255 	md_kva = false;
    256 
    257 #ifdef __HAVE_MM_MD_DIRECT_MAPPED_IO
    258 	paddr_t paddr;
    259 	/* MD case: is this is a directly mapped address? */
    260 	if (mm_md_direct_mapped_io(addr, &paddr)) {
    261 		/* If so, validate physical address. */
    262 		error = mm_md_physacc(paddr, prot);
    263 		if (error) {
    264 			return error;
    265 		}
    266 		md_kva = true;
    267 	}
    268 #endif
    269 	if (!md_kva) {
    270 		bool checked = false;
    271 
    272 #ifdef __HAVE_MM_MD_KERNACC
    273 		/* MD check for the address. */
    274 		error = mm_md_kernacc(addr, prot, &checked);
    275 		if (error) {
    276 			return error;
    277 		}
    278 #endif
    279 		/* UVM check for the address (unless MD indicated to not). */
    280 		if (!checked && !uvm_kernacc(addr, len, prot)) {
    281 			return EFAULT;
    282 		}
    283 	}
    284 	error = uiomove(addr, len, uio);
    285 	return error;
    286 }
    287 
    288 /*
    289  * dev_zero_readwrite: helper for DEV_ZERO (/dev/null) case of R/W.
    290  */
    291 static inline int
    292 dev_zero_readwrite(struct uio *uio, struct iovec *iov)
    293 {
    294 	size_t len;
    295 
    296 	/* Nothing to do for the write case. */
    297 	if (uio->uio_rw == UIO_WRITE) {
    298 		uio->uio_resid = 0;
    299 		return 0;
    300 	}
    301 	/*
    302 	 * Read in page-by-page basis, caller will continue.
    303 	 * Cut appropriately for a single/last-iteration cases.
    304 	 */
    305 	len = MIN(iov->iov_len, PAGE_SIZE);
    306 	return uiomove(dev_zero_page, len, uio);
    307 }
    308 
    309 /*
    310  * mm_readwrite: general memory R/W function.
    311  */
    312 static int
    313 mm_readwrite(dev_t dev, struct uio *uio, int flags)
    314 {
    315 	struct iovec *iov;
    316 	int error;
    317 
    318 #ifdef __HAVE_MM_MD_READWRITE
    319 	/* If defined - there are extra MD cases. */
    320 	switch (minor(dev)) {
    321 	case DEV_MEM:
    322 	case DEV_KMEM:
    323 	case DEV_NULL:
    324 	case DEV_ZERO:
    325 #if defined(COMPAT_16) && defined(__arm)
    326 	case _DEV_ZERO_oARM:
    327 #endif
    328 		break;
    329 	default:
    330 		return mm_md_readwrite(dev, uio);
    331 	}
    332 #endif
    333 	error = 0;
    334 	while (uio->uio_resid > 0 && error == 0) {
    335 		iov = uio->uio_iov;
    336 		if (iov->iov_len == 0) {
    337 			/* Processed; next I/O vector. */
    338 			uio->uio_iov++;
    339 			uio->uio_iovcnt--;
    340 			KASSERT(uio->uio_iovcnt >= 0);
    341 			continue;
    342 		}
    343 		/* Helper functions will process in page-by-page basis. */
    344 		switch (minor(dev)) {
    345 		case DEV_MEM:
    346 			error = dev_mem_readwrite(uio, iov);
    347 			break;
    348 		case DEV_KMEM:
    349 			error = dev_kmem_readwrite(uio, iov);
    350 			break;
    351 		case DEV_NULL:
    352 			if (uio->uio_rw == UIO_WRITE) {
    353 				uio->uio_resid = 0;
    354 			}
    355 			/* Break directly out of the loop. */
    356 			return 0;
    357 		case DEV_FULL:
    358 			if (uio->uio_rw == UIO_WRITE) {
    359 				return ENOSPC;
    360 			}
    361 #if defined(COMPAT_16) && defined(__arm)
    362 			/* FALLTHROUGH */
    363 		case _DEV_ZERO_oARM:
    364 #endif
    365 		/* FALLTHROUGH */
    366 		case DEV_ZERO:
    367 			error = dev_zero_readwrite(uio, iov);
    368 			break;
    369 		default:
    370 			error = ENXIO;
    371 			break;
    372 		}
    373 	}
    374 	return error;
    375 }
    376 
    377 /*
    378  * mm_mmap: general mmap() handler.
    379  */
    380 static paddr_t
    381 mm_mmap(dev_t dev, off_t off, int acc)
    382 {
    383 	vm_prot_t prot;
    384 
    385 #ifdef __HAVE_MM_MD_MMAP
    386 	/* If defined - there are extra mmap() MD cases. */
    387 	switch (minor(dev)) {
    388 	case DEV_MEM:
    389 	case DEV_KMEM:
    390 	case DEV_NULL:
    391 #if defined(COMPAT_16) && defined(__arm)
    392 	case _DEV_ZERO_oARM:
    393 #endif
    394 	case DEV_ZERO:
    395 		break;
    396 	default:
    397 		return mm_md_mmap(dev, off, acc);
    398 	}
    399 #endif
    400 	/*
    401 	 * /dev/null does not make sense, /dev/kmem is volatile and
    402 	 * /dev/zero is handled in mmap already.
    403 	 */
    404 	if (minor(dev) != DEV_MEM) {
    405 		return -1;
    406 	}
    407 
    408 	prot = 0;
    409 	if (acc & PROT_EXEC)
    410 		prot |= VM_PROT_EXECUTE;
    411 	if (acc & PROT_READ)
    412 		prot |= VM_PROT_READ;
    413 	if (acc & PROT_WRITE)
    414 		prot |= VM_PROT_WRITE;
    415 
    416 	/* Validate the physical address. */
    417 	if (mm_md_physacc(off, prot) != 0) {
    418 		return -1;
    419 	}
    420 	return off >> PGSHIFT;
    421 }
    422 
    423 static int
    424 mm_ioctl(dev_t dev, u_long cmd, void *data, int flag, struct lwp *l)
    425 {
    426 
    427 	switch (cmd) {
    428 	case FIONBIO:
    429 		/* We never block anyway. */
    430 		return 0;
    431 
    432 	case FIOSETOWN:
    433 	case FIOGETOWN:
    434 	case TIOCGPGRP:
    435 	case TIOCSPGRP:
    436 	case TIOCGETA:
    437 		return ENOTTY;
    438 
    439 	case FIOASYNC:
    440 		if ((*(int *)data) == 0) {
    441 			return 0;
    442 		}
    443 		/* FALLTHROUGH */
    444 	default:
    445 		return EOPNOTSUPP;
    446 	}
    447 }
    448