io/pci/pci_reloc.c

/*
 * 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 2006 Sun Microsystems, Inc.  All rights reserved.
 * Use is subject to license terms.
 */

#pragma ident   "%Z%%M% %I% %E% SMI"

/*
 * PCI nexus DVMA relocation routines.
 *
 * These routines handle the interactions with the HAT layer to
 * implement page relocation for page(s) which have active DMA handle
 * bindings when DVMA is being used for those handles.
 *
 * The current modus operandi is as follows:
 *
 *   Object binding: register the appropriate callback for each page
 *     of the kernel object while obtaining the PFN for the DVMA page.
 *
 *   Object unbinding: unregister the callback for each page of the
 *     kernel object.
 *
 *   Relocation request:
 *     1) Suspend the bus and sync the caches.
 *     2) Remap the DVMA object using the new provided PFN.
 *     3) Unsuspend the bus.
 *
 *  The relocation code runs with CPUs captured (idling in xc_loop())
 *  so we can only acquire spinlocks at PIL >= 13 for synchronization
 *  within those codepaths.
 */
#include <sys/types.h>
#include <sys/kmem.h>
#include <sys/async.h>
#include <sys/sysmacros.h>
#include <sys/sunddi.h>
#include <sys/machsystm.h>
#include <sys/ddi_impldefs.h>
#include <sys/dvma.h>
#include <vm/hat.h>
#include <sys/pci/pci_obj.h>

/*LINTLIBRARY*/

void
pci_dvma_unregister_callbacks(pci_t *pci_p, ddi_dma_impl_t *mp)
{
    ddi_dma_obj_t *dobj_p = &mp->dmai_object;
    struct as *as_p = dobj_p->dmao_obj.virt_obj.v_as;
    page_t **pplist = dobj_p->dmao_obj.virt_obj.v_priv;
    caddr_t vaddr = dobj_p->dmao_obj.virt_obj.v_addr;
    struct hat *hat_p;
    uint32_t offset;
    int i;

    if (!PCI_DMA_CANRELOC(mp))
        return;

    hat_p = (as_p == NULL)? kas.a_hat : as_p->a_hat;
    ASSERT(hat_p == kas.a_hat);
    ASSERT(pplist == NULL);

    offset = mp->dmai_roffset;
    hat_delete_callback(vaddr, IOMMU_PAGE_SIZE - offset, mp, HAC_PAGELOCK,
        MP_HAT_CB_COOKIE(mp, 0));
    vaddr = (caddr_t)(((uintptr_t)vaddr + IOMMU_PAGE_SIZE) &
        IOMMU_PAGE_MASK);
    for (i = 1; i < mp->dmai_ndvmapages; i++) {
        hat_delete_callback(vaddr, IOMMU_PAGE_SIZE, mp, HAC_PAGELOCK,
            MP_HAT_CB_COOKIE(mp, i));
        vaddr += IOMMU_PAGE_SIZE;
    }
    mp->dmai_flags &= ~DMAI_FLAGS_RELOC;
}

static int
pci_dvma_postrelocator(caddr_t va, uint_t len, uint_t flags, void *mpvoid,
    pfn_t newpfn)
{
    ddi_dma_impl_t *mp = (ddi_dma_impl_t *)mpvoid;
    dev_info_t *rdip = mp->dmai_rdip;
    ddi_dma_obj_t *dobj_p = &mp->dmai_object;
    page_t **pplist = dobj_p->dmao_obj.virt_obj.v_priv;
    caddr_t baseva = dobj_p->dmao_obj.virt_obj.v_addr;
    int index;
    size_t length = IOMMU_PTOB(1);
    off_t offset;

    DEBUG0(DBG_RELOC, rdip, "postrelocator called\n");

    if (flags == HAT_POSTUNSUSPEND) {
        mutex_enter(&pci_reloc_mutex);
        ASSERT(pci_reloc_thread == curthread);
        ASSERT(pci_reloc_presuspend > 0);
        if (--pci_reloc_presuspend == 0) {
            pci_reloc_thread = NULL;
            cv_broadcast(&pci_reloc_cv);
        }
        mutex_exit(&pci_reloc_mutex);
        return (0);
    }

    ASSERT(flags == HAT_UNSUSPEND);
    ASSERT(pci_reloc_suspend > 0);
    pci_reloc_suspend--;

    ASSERT(len <= length);
    ASSERT(pplist == NULL); /* addr bind handle only */
    ASSERT(dobj_p->dmao_obj.virt_obj.v_as == &kas ||
        dobj_p->dmao_obj.virt_obj.v_as == NULL);
    ASSERT(PCI_DMA_ISDVMA(mp));
    ASSERT(pci_reloc_thread == curthread);

    offset = va - baseva;
    index = IOMMU_BTOPR(offset);
    ASSERT(index < mp->dmai_ndvmapages);

    DEBUG3(DBG_RELOC, rdip, "index 0x%x, vaddr 0x%llx, baseva 0x%llx\n",
        index, (int64_t)va, (int64_t)baseva);

    if ((mp)->dmai_ndvmapages == 1) {
        DEBUG2(DBG_RELOC, rdip, "pfn remap (1) 0x%x -> 0x%x\n",
            mp->dmai_pfnlst, newpfn);
            mp->dmai_pfnlst = (void *)newpfn;
    } else {
        DEBUG3(DBG_RELOC, rdip, "pfn remap (%d) 0x%x -> 0x%x\n",
            index, ((iopfn_t *)mp->dmai_pfnlst)[index], newpfn);
        ((iopfn_t *)mp->dmai_pfnlst)[index] = (iopfn_t)newpfn;
    }

    if (ddi_dma_mctl(rdip, rdip, (ddi_dma_handle_t)mp, DDI_DMA_REMAP,
        &offset, &length, NULL, 0) != DDI_SUCCESS)
        return (EIO);
    if (ddi_ctlops(rdip, rdip, DDI_CTLOPS_UNQUIESCE, NULL, NULL) !=
        DDI_SUCCESS)
        return (EIO);

    return (0);
}

/*
 * Log a warning message if a callback is still registered on
 * a page which is being freed.  This is indicative of a driver
 * bug -- DMA handles are bound, and the memory is being freed by
 * the VM subsystem without an unbind call on the handle first.
 */
static int
pci_dma_relocerr(caddr_t va, uint_t len, uint_t errorcode, void *mpvoid)
{
    int errlevel = pci_dma_panic_on_leak? CE_PANIC : CE_WARN;
    if (errorcode == HAT_CB_ERR_LEAKED) {
        cmn_err(errlevel, "object 0x%p has a bound DMA handle 0x%p\n",
            va, mpvoid);
        return (0);
    }

    /* unknown error code, unhandled so panic */
    return (EINVAL);
}

/*
 * pci DVMA remap entry points
 *
 * Called in response to a DDI_DMA_REMAP DMA ctlops command.
 * Remaps the region specified in the underlying IOMMU. Safe
 * to assume that the bus was quiesced and ddi_dma_sync() was
 * invoked by the caller before we got to this point.
 */
int
pci_dvma_remap(dev_info_t *dip, dev_info_t *rdip, ddi_dma_impl_t *mp,
    off_t offset, size_t length)
{
    pci_t *pci_p = get_pci_soft_state(ddi_get_instance(dip));
    iommu_t *iommu_p = pci_p->pci_iommu_p;
    dvma_addr_t dvma_pg;
    size_t npgs;
    int idx;

    dvma_pg = IOMMU_BTOP(mp->dmai_mapping);
    idx = IOMMU_BTOPR(offset);
    dvma_pg += idx;
    npgs = IOMMU_BTOPR(length);

    DEBUG3(DBG_RELOC, mp->dmai_rdip,
        "pci_dvma_remap: dvma_pg 0x%llx len 0x%llx idx 0x%x\n",
        dvma_pg, length, idx);

    ASSERT(pci_p->pci_pbm_p->pbm_quiesce_count > 0);
    iommu_remap_pages(iommu_p, mp, dvma_pg, npgs, idx);

    return (DDI_SUCCESS);
}

void
pci_fdvma_remap(ddi_dma_impl_t *mp, caddr_t kvaddr, dvma_addr_t dvma_pg,
    size_t npages, size_t index, pfn_t newpfn)
{
    fdvma_t *fdvma_p = (fdvma_t *)mp->dmai_fdvma;
    pci_t *pci_p = (pci_t *)fdvma_p->softsp;
    iommu_t *iommu_p = pci_p->pci_iommu_p;
    dev_info_t *dip = pci_p->pci_dip;
    iopfn_t pfn = (iopfn_t)newpfn;
    dvma_addr_t pg_index = dvma_pg - iommu_p->dvma_base_pg;
    int i;
    uint64_t tte;

    /* make sure we don't exceed reserved boundary */
    DEBUG3(DBG_FAST_DVMA, dip, "fast remap index=%x: %p, npgs=%x", index,
        kvaddr, npages);
    if (index + npages > mp->dmai_ndvmapages) {
        cmn_err(pci_panic_on_fatal_errors ? CE_PANIC : CE_WARN,
            "%s%d: fdvma remap index(%lx)+pgs(%lx) exceeds limit\n",
            ddi_driver_name(dip), ddi_get_instance(dip),
            index, npages);
        return;
    }

    for (i = 0; i < npages; i++, kvaddr += IOMMU_PAGE_SIZE) {
        DEBUG3(DBG_FAST_DVMA, dip, "remap dvma_pg %x -> pfn %x,"
            " old tte 0x%llx\n", dvma_pg + i, pfn,
            iommu_p->iommu_tsb_vaddr[pg_index + i]);

        if (pfn == PFN_INVALID)
            goto bad_pfn;

        if (i == 0)
            tte = MAKE_TTE_TEMPLATE(pfn, mp);

        /* XXX assumes iommu and mmu has same page size */
        iommu_p->iommu_tsb_vaddr[pg_index + i] = tte | IOMMU_PTOB(pfn);
        IOMMU_PAGE_FLUSH(iommu_p, (dvma_pg + i));
    }
    return;
bad_pfn:
    cmn_err(CE_WARN, "%s%d: fdvma remap can't get page frame for vaddr %p",
        ddi_driver_name(dip), ddi_get_instance(dip), kvaddr);
}

static int
pci_fdvma_prerelocator(caddr_t va, uint_t len, uint_t flags, void *mpvoid)
{
    ddi_dma_impl_t *mp = (ddi_dma_impl_t *)mpvoid;
    fdvma_t *fdvma_p = (fdvma_t *)mp->dmai_fdvma;
    caddr_t baseva, endva;
    int i;

    /*
     * It isn't safe to do relocation if all of the IOMMU
     * mappings haven't yet been established at this index.
     */
    for (i = 0; i < mp->dmai_ndvmapages; i++) {
        baseva = fdvma_p->kvbase[i];
        endva = baseva + IOMMU_PTOB(fdvma_p->pagecnt[i]);
        if (va >= baseva && va < endva)
            return (0); /* found a valid index */
    }
    return (EAGAIN);
}

static int
pci_fdvma_postrelocator(caddr_t va, uint_t len, uint_t flags, void *mpvoid,
    pfn_t pfn)
{
    ddi_dma_impl_t *mp = (ddi_dma_impl_t *)mpvoid;
    dev_info_t *rdip = mp->dmai_rdip;
    fdvma_t *fdvma_p = (fdvma_t *)mp->dmai_fdvma;
    caddr_t baseva;
    dvma_addr_t dvma_pg;
    size_t length = PAGESIZE;
    int i;

    DEBUG0(DBG_RELOC, rdip, "fdvma postrelocator called\n");

    if (flags == HAT_POSTUNSUSPEND) {
        mutex_enter(&pci_reloc_mutex);
        ASSERT(pci_reloc_thread == curthread);
        if (--pci_reloc_presuspend == 0) {
            pci_reloc_thread = NULL;
            cv_broadcast(&pci_reloc_cv);
        }
        mutex_exit(&pci_reloc_mutex);
        return (0);
    }

    pci_reloc_suspend--;

    ASSERT(flags == HAT_UNSUSPEND);
    ASSERT(len <= length);
    ASSERT((mp->dmai_rflags & DMP_BYPASSNEXUS) != 0);

    /*
     * This virtual page can have multiple cookies that refer
     * to it within the same handle. We must walk the whole
     * table for this DMA handle finding all the cookies, and
     * update all of them. Sigh.
     */
    for (i = 0; i < mp->dmai_ndvmapages; i++) {
        caddr_t endva;
        int index;

        baseva = fdvma_p->kvbase[i];
        endva = baseva + IOMMU_PTOB(fdvma_p->pagecnt[i]);

        if (va >= baseva && va < endva) {
            index = i + IOMMU_BTOP(va - baseva);
            ASSERT(index < mp->dmai_ndvmapages);

            DEBUG4(DBG_RELOC, rdip, "mp %p: index 0x%x, "
                " vaddr 0x%llx, baseva 0x%llx\n", mp, index,
                (int64_t)va, (int64_t)baseva);

            dvma_pg = IOMMU_BTOP(mp->dmai_mapping) + index;
            pci_fdvma_remap(mp, va, dvma_pg, IOMMU_BTOP(length),
                index, pfn);
        }
    }

    if (ddi_ctlops(rdip, rdip, DDI_CTLOPS_UNQUIESCE, NULL, NULL) !=
        DDI_SUCCESS)
        return (EIO);

    return (0);
}

void
pci_fdvma_unregister_callbacks(pci_t *pci_p, fdvma_t *fdvma_p,
    ddi_dma_impl_t *mp, uint_t index)
{
    size_t npgs = fdvma_p->pagecnt[index];
    caddr_t kva = fdvma_p->kvbase[index];
    int i;

    ASSERT(index + npgs <= mp->dmai_ndvmapages);
    ASSERT(kva != NULL);

    for (i = 0; i < npgs && pci_dvma_remap_enabled;
        i++, kva += IOMMU_PAGE_SIZE)
        hat_delete_callback(kva, IOMMU_PAGE_SIZE, mp, HAC_PAGELOCK,
            fdvma_p->cbcookie[index + i]);
}

static int
pci_common_prerelocator(caddr_t va, uint_t len, uint_t flags, void *mpvoid)
{
    ddi_dma_impl_t *mp = (ddi_dma_impl_t *)mpvoid;
    ddi_dma_handle_t h = (ddi_dma_handle_t)mpvoid;
    dev_info_t *rdip = mp->dmai_rdip;
    int ret;

    DEBUG0(DBG_RELOC, rdip, "prerelocator called\n");

    if (flags == HAT_PRESUSPEND) {
        if (!ddi_prop_exists(DDI_DEV_T_ANY, rdip, DDI_PROP_NOTPROM,
            "dvma-remap-supported"))
            return (ENOTSUP);
        if (!PCI_DMA_ISMAPPED(mp))
            return (EAGAIN);

        if (mp->dmai_rflags & DMP_BYPASSNEXUS) {
            ret = pci_fdvma_prerelocator(va, len, flags, mpvoid);
            if (ret != 0)
                return (ret);
        } else if (!PCI_DMA_ISDVMA(mp))
            return (EINVAL);

        /*
         * Acquire the exclusive right to relocate a PCI DMA page,
         * since we later have to pause CPUs which could otherwise
         * lead to all sorts of synchronization headaches.
         */
        mutex_enter(&pci_reloc_mutex);
        if (pci_reloc_thread != curthread) {
            while (pci_reloc_thread != NULL) {
                cv_wait(&pci_reloc_cv, &pci_reloc_mutex);
            }
            pci_reloc_thread = curthread;
            ASSERT(pci_reloc_suspend == 0);
        }
        mutex_exit(&pci_reloc_mutex);

        ASSERT(pci_reloc_thread == curthread);
        pci_reloc_presuspend++;

        return (0);
    }

    ASSERT(flags == HAT_SUSPEND);
    ASSERT(PCI_DMA_CANRELOC(mp));
    ASSERT(pci_reloc_thread == curthread);
    pci_reloc_suspend++;

    if (ddi_ctlops(rdip, rdip, DDI_CTLOPS_QUIESCE, NULL, NULL) !=
        DDI_SUCCESS)
        return (EIO);
    if (ddi_dma_sync(h, 0, 0, DDI_DMA_SYNC_FORKERNEL) != DDI_SUCCESS)
        return (EIO);

    return (0);
}

/*
 * Register two callback types: one for normal DVMA and the
 * other for fast DVMA, since each method has a different way
 * of tracking the PFNs behind a handle.
 */
void
pci_reloc_init(void)
{
    int key = pci_reloc_getkey();

    mutex_init(&pci_reloc_mutex, NULL, MUTEX_DEFAULT, NULL);
    cv_init(&pci_reloc_cv, NULL, CV_DEFAULT, NULL);
    pci_dvma_cbid = hat_register_callback(
        key + ('D'<<24 | 'V'<<16 | 'M'<<8 | 'A'),
        pci_common_prerelocator, pci_dvma_postrelocator,
        pci_dma_relocerr, 1);
    pci_fast_dvma_cbid = hat_register_callback(
        key + ('F'<<24 | 'D'<<16 | 'M'<<8 | 'A'),
        pci_common_prerelocator,
        pci_fdvma_postrelocator, pci_dma_relocerr, 1);
}

void
pci_reloc_fini(void)
{
    cv_destroy(&pci_reloc_cv);
    mutex_destroy(&pci_reloc_mutex);
}