[src/trunk]: src/sys/arch/xen/xen revert previous - physio arranges for tranf...

[jdolecek <jdolecek%NetBSD.org@localhost>]

details:   https://anonhg.NetBSD.org/src/rev/d0e407a89f21
branches:  trunk
changeset: 1009068:d0e407a89f21
user:      jdolecek <jdolecek%NetBSD.org@localhost>
date:      Sat Apr 11 17:52:01 2020 +0000

description:
revert previous - physio arranges for tranfer directly to user-provided
buffers, which are generally not DEV_BSIZE-aligned

diffstat:

 sys/arch/xen/xen/xbd_xenbus.c |  69 +++++++++++++++++++++++++++++++++++-------
 1 files changed, 57 insertions(+), 12 deletions(-)

diffs (158 lines):

diff -r 2530f5e16b30 -r d0e407a89f21 sys/arch/xen/xen/xbd_xenbus.c
--- a/sys/arch/xen/xen/xbd_xenbus.c     Sat Apr 11 17:43:54 2020 +0000
+++ b/sys/arch/xen/xen/xbd_xenbus.c     Sat Apr 11 17:52:01 2020 +0000
@@ -1,4 +1,4 @@
-/*      $NetBSD: xbd_xenbus.c,v 1.102 2020/04/11 16:15:34 jdolecek Exp $      */
+/*      $NetBSD: xbd_xenbus.c,v 1.103 2020/04/11 17:52:01 jdolecek Exp $      */
 
 /*
  * Copyright (c) 2006 Manuel Bouyer.
@@ -50,7 +50,7 @@
  */
 
 #include <sys/cdefs.h>
-__KERNEL_RCSID(0, "$NetBSD: xbd_xenbus.c,v 1.102 2020/04/11 16:15:34 jdolecek Exp $");
+__KERNEL_RCSID(0, "$NetBSD: xbd_xenbus.c,v 1.103 2020/04/11 17:52:01 jdolecek Exp $");
 
 #include "opt_xen.h"
 
@@ -97,8 +97,6 @@
 #define XEN_BSHIFT      9               /* log2(XEN_BSIZE) */
 #define XEN_BSIZE       (1 << XEN_BSHIFT) 
 
-__CTASSERT((DEV_BSIZE == XEN_BSIZE));
-
 struct xbd_req {
        SLIST_ENTRY(xbd_req) req_next;
        uint16_t req_id; /* ID passed to backend */
@@ -107,6 +105,7 @@
                grant_ref_t req_gntref[BLKIF_MAX_SEGMENTS_PER_REQUEST];
                int req_nr_segments; /* number of segments in this request */
                struct buf *req_bp; /* buffer associated with this request */
+               void *req_data; /* pointer to the data buffer */
            } req_rw;
            struct {
                int s_error;
@@ -117,6 +116,7 @@
 #define req_gntref     u.req_rw.req_gntref
 #define req_nr_segments        u.req_rw.req_nr_segments
 #define req_bp         u.req_rw.req_bp
+#define req_data       u.req_rw.req_data
 #define req_sync       u.req_sync
 
 struct xbd_xenbus_softc {
@@ -156,6 +156,7 @@
 #define BLKIF_FEATURE_PERSISTENT       0x4
 #define BLKIF_FEATURE_BITS             \
        "\20\1CACHE-FLUSH\2BARRIER\3PERSISTENT"
+       struct evcnt sc_cnt_map_unalign;
 };
 
 #if 0
@@ -176,6 +177,9 @@
 static void xbd_backend_changed(void *, XenbusState);
 static void xbd_connect(struct xbd_xenbus_softc *);
 
+static int  xbd_map_align(struct xbd_req *);
+static void xbd_unmap_align(struct xbd_req *);
+
 static void xbdminphys(struct buf *);
 
 CFATTACH_DECL3_NEW(xbd, sizeof(struct xbd_xenbus_softc),
@@ -281,6 +285,9 @@
                panic("%s: can't alloc ring", device_xname(self));
        sc->sc_ring.sring = ring;
 
+       evcnt_attach_dynamic(&sc->sc_cnt_map_unalign, EVCNT_TYPE_MISC,
+           NULL, device_xname(self), "map unaligned");
+
        /* resume shared structures and tell backend that we are ready */
        if (xbd_xenbus_resume(self, PMF_Q_NONE) == false) {
                uvm_km_free(kernel_map, (vaddr_t)ring, PAGE_SIZE,
@@ -364,6 +371,8 @@
        uvm_km_free(kernel_map, (vaddr_t)sc->sc_ring.sring,
            PAGE_SIZE, UVM_KMF_WIRED);
 
+       evcnt_detach(&sc->sc_cnt_map_unalign);
+
        pmf_device_deregister(dev);
 
        return 0;
@@ -702,6 +711,9 @@
                        goto next;
                }
                /* b_resid was set in dk_start */
+               if (__predict_false(
+                   xbdreq->req_data != NULL && bp->b_data != xbdreq->req_data))
+                       xbd_unmap_align(xbdreq);
 next:
                xbdreq->req_bp = NULL;
                dk_done(&sc->sc_dksc, bp);
@@ -971,13 +983,16 @@
        }
 
        xbdreq->req_bp = bp;
+       xbdreq->req_data = bp->b_data;
+       if (__predict_false((vaddr_t)bp->b_data & (XEN_BSIZE - 1))) {
+               sc->sc_cnt_map_unalign.ev_count++;
 
-       /*
-        * All bufs passed by system are aligned to DEV_BSIZE.
-        * xbd requires this to be the case, as transfer offsets
-        * are expressed in multiplies of 512 (XEN_BSIZE).
-        */
-       KASSERT(((vaddr_t)bp->b_data & (XEN_BSIZE - 1)) == 0);
+               if (__predict_false(xbd_map_align(xbdreq) != 0)) {
+                       DPRINTF(("xbd_diskstart: no align\n"));
+                       error = EAGAIN;
+                       goto out;
+               }
+       }
 
        SLIST_REMOVE_HEAD(&sc->sc_xbdreq_head, req_next);
        req = RING_GET_REQUEST(&sc->sc_ring, sc->sc_ring.req_prod_pvt);
@@ -987,8 +1002,8 @@
        req->sector_number = bp->b_rawblkno;
        req->handle = sc->sc_handle;
 
-       va = (vaddr_t)bp->b_data & ~PAGE_MASK;
-       off = (vaddr_t)bp->b_data & PAGE_MASK;
+       va = (vaddr_t)xbdreq->req_data & ~PAGE_MASK;
+       off = (vaddr_t)xbdreq->req_data & PAGE_MASK;
        bcount = bp->b_bcount;
        bp->b_resid = 0;
        for (seg = 0; bcount > 0;) {
@@ -1028,3 +1043,33 @@
 err:
        return error;
 }
+
+static int
+xbd_map_align(struct xbd_req *req)
+{
+       int s = splvm(); /* XXXSMP - bogus? */
+       int rc;
+
+       rc = uvm_km_kmem_alloc(kmem_va_arena,
+           req->req_bp->b_bcount, (VM_NOSLEEP | VM_INSTANTFIT),
+           (vmem_addr_t *)&req->req_data);
+       splx(s);
+       if (__predict_false(rc != 0))
+               return ENOMEM;
+       if ((req->req_bp->b_flags & B_READ) == 0)
+               memcpy(req->req_data, req->req_bp->b_data,
+                   req->req_bp->b_bcount);
+       return 0;
+}
+
+static void
+xbd_unmap_align(struct xbd_req *req)
+{
+       int s;
+       if (req->req_bp->b_flags & B_READ)
+               memcpy(req->req_bp->b_data, req->req_data,
+                   req->req_bp->b_bcount);
+       s = splvm(); /* XXXSMP - bogus? */
+       uvm_km_kmem_free(kmem_va_arena, (vaddr_t)req->req_data, req->req_bp->b_bcount);
+       splx(s);
+}