root/sys/arch/alpha/pci/cia_dma.c
/* $NetBSD: cia_dma.c,v 1.38 2024/02/09 22:08:31 andvar Exp $ */

/*-
 * Copyright (c) 1997, 1998 The NetBSD Foundation, Inc.
 * All rights reserved.
 *
 * This code is derived from software contributed to The NetBSD Foundation
 * by Jason R. Thorpe of the Numerical Aerospace Simulation Facility,
 * NASA Ames Research Center.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions
 * are met:
 * 1. Redistributions of source code must retain the above copyright
 *    notice, this list of conditions and the following disclaimer.
 * 2. Redistributions in binary form must reproduce the above copyright
 *    notice, this list of conditions and the following disclaimer in the
 *    documentation and/or other materials provided with the distribution.
 *
 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
 * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
 * POSSIBILITY OF SUCH DAMAGE.
 */

#include <sys/cdefs.h>                  /* RCS ID & Copyright macro defns */

__KERNEL_RCSID(0, "$NetBSD: cia_dma.c,v 1.38 2024/02/09 22:08:31 andvar Exp $");

#include <sys/param.h>
#include <sys/systm.h>
#include <sys/kernel.h>
#include <sys/device.h>

#define _ALPHA_BUS_DMA_PRIVATE
#include <sys/bus.h>

#include <uvm/uvm_extern.h>

#include <dev/pci/pcireg.h>
#include <dev/pci/pcivar.h>
#include <alpha/pci/ciareg.h>
#include <alpha/pci/ciavar.h>

static bus_dma_tag_t cia_dma_get_tag(bus_dma_tag_t, alpha_bus_t);

static int      cia_bus_dmamap_create_direct(bus_dma_tag_t, bus_size_t, int,
                    bus_size_t, bus_size_t, int, bus_dmamap_t *);

static int      cia_bus_dmamap_load_sgmap(bus_dma_tag_t, bus_dmamap_t, void *,
                    bus_size_t, struct proc *, int);

static int      cia_bus_dmamap_load_mbuf_sgmap(bus_dma_tag_t, bus_dmamap_t,
                    struct mbuf *, int);

static int      cia_bus_dmamap_load_uio_sgmap(bus_dma_tag_t, bus_dmamap_t,
                    struct uio *, int);

static int      cia_bus_dmamap_load_raw_sgmap(bus_dma_tag_t, bus_dmamap_t,
                    bus_dma_segment_t *, int, bus_size_t, int);

static void     cia_bus_dmamap_unload_sgmap(bus_dma_tag_t, bus_dmamap_t);

/*
 * Direct-mapped window: 1G at 1G
 */
#define CIA_DIRECT_MAPPED_BASE  (1UL*1024*1024*1024)
#define CIA_DIRECT_MAPPED_SIZE  (1UL*1024*1024*1024)

/*
 * SGMAP window for ISA: 8M at 8M
 */
#define CIA_SGMAP_MAPPED_LO_BASE (8UL*1024*1024)
#define CIA_SGMAP_MAPPED_LO_SIZE (8UL*1024*1024)

/*
 * SGMAP window for PCI: 1G at 3G
 */
#define CIA_SGMAP_MAPPED_HI_BASE (3UL*1024*1024*1024)
#define CIA_SGMAP_MAPPED_HI_SIZE (1UL*1024*1024*1024)

/* ALCOR/ALGOR2/PYXIS have a 256-byte out-bound DMA prefetch threshold. */
#define CIA_SGMAP_PFTHRESH      256

static void     cia_tlb_invalidate(void);
static void     cia_broken_pyxis_tlb_invalidate(void);

static void     (*cia_tlb_invalidate_fn)(void);

#define CIA_TLB_INVALIDATE()    (*cia_tlb_invalidate_fn)()

struct alpha_sgmap cia_pyxis_bug_sgmap;
#define CIA_PYXIS_BUG_BASE      (128UL*1024*1024)
#define CIA_PYXIS_BUG_SIZE      (2UL*1024*1024)

static void
cia_dma_shutdown(void *arg)
{
        struct cia_config *ccp = arg;
        int i;

        /*
         * Restore the original values, to make the firmware happy.
         */
        for (i = 0; i < 4; i++) {
                REGVAL(CIA_PCI_W0BASE + (i * 0x100)) =
                    ccp->cc_saved_windows.wbase[i];
                alpha_mb();
                REGVAL(CIA_PCI_W0MASK + (i * 0x100)) =
                    ccp->cc_saved_windows.wmask[i];
                alpha_mb();
                REGVAL(CIA_PCI_T0BASE + (i * 0x100)) =
                    ccp->cc_saved_windows.tbase[i];
                alpha_mb();
        }
}

void
cia_dma_init(struct cia_config *ccp)
{
        bus_addr_t tbase;
        bus_dma_tag_t t;
        bus_dma_tag_t t_sg_hi = NULL;
        int i;

        /*
         * Save our configuration to restore at shutdown, just
         * in case the firmware would get cranky with us.
         */
        for (i = 0; i < 4; i++) {
                ccp->cc_saved_windows.wbase[i] =
                    REGVAL(CIA_PCI_W0BASE + (i * 0x100));
                ccp->cc_saved_windows.wmask[i] =
                    REGVAL(CIA_PCI_W0MASK + (i * 0x100));
                ccp->cc_saved_windows.tbase[i] =
                    REGVAL(CIA_PCI_T0BASE + (i * 0x100));
        }
        shutdownhook_establish(cia_dma_shutdown, ccp);

        /*
         * If we have more than 1GB of RAM, then set up an sgmap-mapped
         * DMA window for PCI.  This is better than using the ISA window,
         * which is pretty small and PCI devices could starve it.
         *
         * N.B. avail_end is "last-usable PFN + 1".
         */
        if (uvm_physseg_get_avail_end(uvm_physseg_get_last()) >
            atop(CIA_DIRECT_MAPPED_SIZE)) {
                t = t_sg_hi = &ccp->cc_dmat_sgmap_hi;
                t->_cookie = ccp;
                t->_wbase = CIA_SGMAP_MAPPED_HI_BASE;
                t->_wsize = CIA_SGMAP_MAPPED_HI_SIZE;
                t->_next_window = NULL;
                t->_boundary = 0;
                t->_sgmap = &ccp->cc_sgmap_hi;
                t->_pfthresh = CIA_SGMAP_PFTHRESH;
                t->_get_tag = cia_dma_get_tag;
                t->_dmamap_create = alpha_sgmap_dmamap_create;
                t->_dmamap_destroy = alpha_sgmap_dmamap_destroy;
                t->_dmamap_load = cia_bus_dmamap_load_sgmap;
                t->_dmamap_load_mbuf = cia_bus_dmamap_load_mbuf_sgmap;
                t->_dmamap_load_uio = cia_bus_dmamap_load_uio_sgmap;
                t->_dmamap_load_raw = cia_bus_dmamap_load_raw_sgmap;
                t->_dmamap_unload = cia_bus_dmamap_unload_sgmap;
                t->_dmamap_sync = _bus_dmamap_sync;

                t->_dmamem_alloc = _bus_dmamem_alloc;
                t->_dmamem_free = _bus_dmamem_free;
                t->_dmamem_map = _bus_dmamem_map;
                t->_dmamem_unmap = _bus_dmamem_unmap;
                t->_dmamem_mmap = _bus_dmamem_mmap;
        }

        /*
         * Initialize the DMA tag used for direct-mapped DMA.
         */
        t = &ccp->cc_dmat_direct;
        t->_cookie = ccp;
        t->_wbase = CIA_DIRECT_MAPPED_BASE;
        t->_wsize = CIA_DIRECT_MAPPED_SIZE;
        t->_next_window = t_sg_hi;
        t->_boundary = 0;
        t->_sgmap = NULL;
        t->_get_tag = cia_dma_get_tag;
        t->_dmamap_create = cia_bus_dmamap_create_direct;
        t->_dmamap_destroy = _bus_dmamap_destroy;
        t->_dmamap_load = _bus_dmamap_load_direct;
        t->_dmamap_load_mbuf = _bus_dmamap_load_mbuf_direct;
        t->_dmamap_load_uio = _bus_dmamap_load_uio_direct;
        t->_dmamap_load_raw = _bus_dmamap_load_raw_direct;
        t->_dmamap_unload = _bus_dmamap_unload;
        t->_dmamap_sync = _bus_dmamap_sync;

        t->_dmamem_alloc = _bus_dmamem_alloc;
        t->_dmamem_free = _bus_dmamem_free;
        t->_dmamem_map = _bus_dmamem_map;
        t->_dmamem_unmap = _bus_dmamem_unmap;
        t->_dmamem_mmap = _bus_dmamem_mmap;

        /*
         * Initialize the DMA tag used for sgmap-mapped ISA DMA.
         */
        t = &ccp->cc_dmat_sgmap_lo;
        t->_cookie = ccp;
        t->_wbase = CIA_SGMAP_MAPPED_LO_BASE;
        t->_wsize = CIA_SGMAP_MAPPED_LO_SIZE;
        t->_next_window = NULL;
        t->_boundary = 0;
        t->_sgmap = &ccp->cc_sgmap_lo;
        t->_pfthresh = CIA_SGMAP_PFTHRESH;
        t->_get_tag = cia_dma_get_tag;
        t->_dmamap_create = alpha_sgmap_dmamap_create;
        t->_dmamap_destroy = alpha_sgmap_dmamap_destroy;
        t->_dmamap_load = cia_bus_dmamap_load_sgmap;
        t->_dmamap_load_mbuf = cia_bus_dmamap_load_mbuf_sgmap;
        t->_dmamap_load_uio = cia_bus_dmamap_load_uio_sgmap;
        t->_dmamap_load_raw = cia_bus_dmamap_load_raw_sgmap;
        t->_dmamap_unload = cia_bus_dmamap_unload_sgmap;
        t->_dmamap_sync = _bus_dmamap_sync;

        t->_dmamem_alloc = _bus_dmamem_alloc;
        t->_dmamem_free = _bus_dmamem_free;
        t->_dmamem_map = _bus_dmamem_map;
        t->_dmamem_unmap = _bus_dmamem_unmap;
        t->_dmamem_mmap = _bus_dmamem_mmap;

        /*
         * The firmware will have set up window 1 as a 1G dirct-mapped
         * DMA window beginning at 1G.  While it's pretty safe to assume
         * this is the case, we'll go ahead and program the registers
         * as we expect as a belt-and-suspenders measure.
         */
        REGVAL(CIA_PCI_W1BASE) = CIA_DIRECT_MAPPED_BASE | CIA_PCI_WnBASE_W_EN;
        alpha_mb();
        REGVAL(CIA_PCI_W1MASK) = CIA_PCI_WnMASK_1G;
        alpha_mb();
        REGVAL(CIA_PCI_T1BASE) = 0;
        alpha_mb();

        /*
         * Initialize the SGMAP(s).  Must align page table to at least 32k
         * (hardware bug?).
         */
        alpha_sgmap_init(t, &ccp->cc_sgmap_lo, "cia_sgmap_lo",
            CIA_SGMAP_MAPPED_LO_BASE, 0, CIA_SGMAP_MAPPED_LO_SIZE,
            sizeof(uint64_t), NULL, (32*1024));
        if (t_sg_hi != NULL) {
                alpha_sgmap_init(t_sg_hi, &ccp->cc_sgmap_hi, "cia_sgmap_hi",
                    CIA_SGMAP_MAPPED_HI_BASE, 0, CIA_SGMAP_MAPPED_HI_SIZE,
                    sizeof(uint64_t), NULL, (32*1024));
        }

        /*
         * Set up window 0 as an 8MB SGMAP-mapped window
         * starting at 8MB.
         */
        REGVAL(CIA_PCI_W0BASE) = CIA_SGMAP_MAPPED_LO_BASE |
            CIA_PCI_WnBASE_SG_EN | CIA_PCI_WnBASE_W_EN;
        alpha_mb();

        REGVAL(CIA_PCI_W0MASK) = CIA_PCI_WnMASK_8M;
        alpha_mb();

        tbase = ccp->cc_sgmap_lo.aps_ptpa >> CIA_PCI_TnBASE_SHIFT;
        if ((tbase & CIA_PCI_TnBASE_MASK) != tbase)
                panic("cia_dma_init: bad page table address");
        REGVAL(CIA_PCI_T0BASE) = tbase;
        alpha_mb();

        /*
         * (Maybe) set up window 3 as a 1G SGMAP-mapped window starting
         * at 3G.
         */
        if (t_sg_hi != NULL) {
                REGVAL(CIA_PCI_W3BASE) = CIA_SGMAP_MAPPED_HI_BASE |
                    CIA_PCI_WnBASE_SG_EN | CIA_PCI_WnBASE_W_EN;
                alpha_mb();

                REGVAL(CIA_PCI_W3MASK) = CIA_PCI_WnMASK_1G;
                alpha_mb();

                tbase = ccp->cc_sgmap_hi.aps_ptpa >> CIA_PCI_TnBASE_SHIFT;
                if ((tbase & CIA_PCI_TnBASE_MASK) != tbase)
                        panic("cia_dma_init: bad page table address");
                REGVAL(CIA_PCI_T3BASE) = tbase;
                alpha_mb();
        } else {
                REGVAL(CIA_PCI_W3BASE) = 0;
                alpha_mb();
        }

        /*
         * Pass 1 and 2 (i.e. revision <= 1) of the Pyxis have a
         * broken scatter/gather TLB; it cannot be invalidated.  To
         * work around this problem, we configure window 2 as an SG
         * 2M window at 128M, which we use in DMA loopback mode to
         * read a spill page.  This works by causing TLB misses,
         * causing the old entries to be purged to make room for
         * the new entries coming in for the spill page.
         */
        if ((ccp->cc_flags & CCF_ISPYXIS) != 0 && ccp->cc_rev <= 1) {
                uint64_t *page_table;

                cia_tlb_invalidate_fn =
                    cia_broken_pyxis_tlb_invalidate;

                alpha_sgmap_init(t, &cia_pyxis_bug_sgmap,
                    "pyxis_bug_sgmap", CIA_PYXIS_BUG_BASE, 0,
                    CIA_PYXIS_BUG_SIZE, sizeof(uint64_t), NULL,
                    (32*1024));

                REGVAL(CIA_PCI_W2BASE) = CIA_PYXIS_BUG_BASE |
                    CIA_PCI_WnBASE_SG_EN | CIA_PCI_WnBASE_W_EN;
                alpha_mb();

                REGVAL(CIA_PCI_W2MASK) = CIA_PCI_WnMASK_2M;
                alpha_mb();

                tbase = cia_pyxis_bug_sgmap.aps_ptpa >>
                    CIA_PCI_TnBASE_SHIFT;
                if ((tbase & CIA_PCI_TnBASE_MASK) != tbase)
                        panic("cia_dma_init: bad page table address");
                REGVAL(CIA_PCI_T2BASE) = tbase;
                alpha_mb();

                /*
                 * Initialize the page table to point at the spill
                 * page.  Leave the last entry invalid.
                 */
                pci_sgmap_pte64_init_spill_page_pte();
                for (i = 0, page_table = cia_pyxis_bug_sgmap.aps_pt;
                     i < (CIA_PYXIS_BUG_SIZE / PAGE_SIZE) - 1; i++) {
                        page_table[i] =
                            pci_sgmap_pte64_prefetch_spill_page_pte;
                }
                alpha_mb();
        } else {
                REGVAL(CIA_PCI_W2BASE) = 0;
                alpha_mb();

                cia_tlb_invalidate_fn = cia_tlb_invalidate;
        }

        CIA_TLB_INVALIDATE();
}

/*
 * Return the bus dma tag to be used for the specified bus type.
 * INTERNAL USE ONLY!
 */
static bus_dma_tag_t
cia_dma_get_tag(bus_dma_tag_t t, alpha_bus_t bustype)
{
        struct cia_config *ccp = t->_cookie;

        switch (bustype) {
        case ALPHA_BUS_PCI:
        case ALPHA_BUS_EISA:
                /*
                 * Regardless if how much memory is installed,
                 * start with the direct-mapped window.  It will
                 * fall back to the SGMAP window if we encounter a
                 * page that is out of range.
                 */
                return (&ccp->cc_dmat_direct);

        case ALPHA_BUS_ISA:
                /*
                 * ISA doesn't have enough address bits to use
                 * the direct-mapped DMA window, so we must use
                 * SGMAPs.
                 */
                return (&ccp->cc_dmat_sgmap_lo);

        default:
                panic("cia_dma_get_tag: shouldn't be here, really...");
        }
}

/*
 * Create a CIA direct-mapped DMA map.
 */
static int
cia_bus_dmamap_create_direct(
        bus_dma_tag_t t,
        bus_size_t size,
        int nsegments,
        bus_size_t maxsegsz,
        bus_size_t boundary,
        int flags,
        bus_dmamap_t *dmamp)
{
        struct cia_config *ccp = t->_cookie;
        bus_dmamap_t map;
        int error;

        error = _bus_dmamap_create(t, size, nsegments, maxsegsz,
            boundary, flags, dmamp);
        if (error)
                return (error);

        map = *dmamp;

        if ((ccp->cc_flags & CCF_PYXISBUG) != 0 &&
            map->_dm_segcnt > 1) {
                /*
                 * We have a Pyxis with the DMA page crossing bug, make
                 * sure we don't coalesce adjacent DMA segments.
                 *
                 * NOTE: We can only do this if the max segment count
                 * is greater than 1.  This is because many network
                 * drivers allocate large contiguous blocks of memory
                 * for control data structures, even though they won't
                 * do any single DMA that crosses a page boundary.
                 *      -- thorpej@NetBSD.org, 2/5/2000
                 */
                map->_dm_flags |= DMAMAP_NO_COALESCE;
        }

        return (0);
}

/*
 * Load a CIA SGMAP-mapped DMA map with a linear buffer.
 */
static int
cia_bus_dmamap_load_sgmap(bus_dma_tag_t t, bus_dmamap_t map, void *buf,
    bus_size_t buflen, struct proc *p, int flags)
{
        int error;

        error = pci_sgmap_pte64_load(t, map, buf, buflen, p, flags,
            t->_sgmap);
        if (error == 0)
                CIA_TLB_INVALIDATE();

        return (error);
}

/*
 * Load a CIA SGMAP-mapped DMA map with an mbuf chain.
 */
static int
cia_bus_dmamap_load_mbuf_sgmap(bus_dma_tag_t t, bus_dmamap_t map,
    struct mbuf *m, int flags)
{
        int error;

        error = pci_sgmap_pte64_load_mbuf(t, map, m, flags, t->_sgmap);
        if (error == 0)
                CIA_TLB_INVALIDATE();

        return (error);
}

/*
 * Load a CIA SGMAP-mapped DMA map with a uio.
 */
static int
cia_bus_dmamap_load_uio_sgmap(bus_dma_tag_t t, bus_dmamap_t map,
    struct uio *uio, int flags)
{
        int error;

        error = pci_sgmap_pte64_load_uio(t, map, uio, flags, t->_sgmap);
        if (error == 0)
                CIA_TLB_INVALIDATE();

        return (error);
}

/*
 * Load a CIA SGMAP-mapped DMA map with raw memory.
 */
static int
cia_bus_dmamap_load_raw_sgmap(bus_dma_tag_t t, bus_dmamap_t map,
    bus_dma_segment_t *segs, int nsegs, bus_size_t size, int flags)
{
        int error;

        error = pci_sgmap_pte64_load_raw(t, map, segs, nsegs, size, flags,
            t->_sgmap);
        if (error == 0)
                CIA_TLB_INVALIDATE();

        return (error);
}

/*
 * Unload a CIA DMA map.
 */
static void
cia_bus_dmamap_unload_sgmap(bus_dma_tag_t t, bus_dmamap_t map)
{

        /*
         * Invalidate any SGMAP page table entries used by this
         * mapping.
         */
        pci_sgmap_pte64_unload(t, map, t->_sgmap);
        CIA_TLB_INVALIDATE();

        /*
         * Do the generic bits of the unload.
         */
        _bus_dmamap_unload_common(t, map);
}

/*
 * Flush the CIA scatter/gather TLB.
 */
static void
cia_tlb_invalidate(void)
{

        alpha_mb();
        REGVAL(CIA_PCI_TBIA) = CIA_PCI_TBIA_ALL;
        alpha_mb();
}

/*
 * Flush the scatter/gather TLB on broken Pyxis chips.
 */
static void
cia_broken_pyxis_tlb_invalidate(void)
{
        uint32_t ctrl;
        int i, s;

        s = splhigh();

        /*
         * Put the Pyxis into PCI loopback mode.
         */
        alpha_mb();
        ctrl = REGVAL(CIA_CSR_CTRL);
        REGVAL(CIA_CSR_CTRL) = ctrl | CTRL_PCI_LOOP_EN;
        alpha_mb();

        /*
         * Now, read from PCI dense memory space at offset 128M (our
         * target window base), skipping 64k on each read.  This forces
         * S/G TLB misses.
         *
         * XXX Looks like the TLB entries are `not quite LRU'.  We need
         * XXX to read more times than there are actual tags!
         */
        for (i = 0; i < CIA_TLB_NTAGS + 4; i++) {
                volatile uint64_t dummy;
                dummy = *((volatile uint64_t *)
                    ALPHA_PHYS_TO_K0SEG(CIA_PCI_DENSE + CIA_PYXIS_BUG_BASE +
                    (i * 65536)));
                __USE(dummy);
        }

        /*
         * Restore normal PCI operation.
         */
        alpha_mb();
        REGVAL(CIA_CSR_CTRL) = ctrl;
        alpha_mb();

        splx(s);
}