[Qemu-devel] [PATCH 14/15] PPC: dbdma: Support unaligned DMA access

[Alexander Graf <agraf@suse.de>]

The DBDMA engine really just reads bytes from a producing device (IDE
in our case) and shoves these bytes into memory. It doesn't care whether
any alignment takes place or not.

Our code today however assumes that block accesses always happen on
sector (512 byte) boundaries. This is a fair assumption for most cases.

However, Mac OS X really likes to do unaligned, incomplete accesses
that it finishes with the next DMA request.

So we need to read / write the unaligned bits independent of the actual
asynchronous request, because that one can only handle 512-byte-aligned
data. We also need to cache these unaligned sectors until the next DMA
request, at which point the data might be successfully flushed from the
pipe.

Signed-off-by: Alexander Graf <agraf@suse.de>
---
 hw/ide/macio.c             | 128 ++++++++++++++++++++++++++++++++++++++++++---
 include/hw/ppc/mac_dbdma.h |   3 ++
 2 files changed, 124 insertions(+), 7 deletions(-)

diff --git a/hw/ide/macio.c b/hw/ide/macio.c
index 095e1f1..202df06 100644
--- a/hw/ide/macio.c
+++ b/hw/ide/macio.c
@@ -51,11 +51,13 @@ static void pmac_ide_atapi_transfer_cb(void *opaque, int ret)
     DBDMA_io *io = opaque;
     MACIOIDEState *m = io->opaque;
     IDEState *s = idebus_active_if(&m->bus);
+    int unaligned;
 
     if (ret < 0) {
         m->aiocb = NULL;
         qemu_sglist_destroy(&s->sg);
         ide_atapi_io_error(s, ret);
+        io->remainder_len = 0;
         goto done;
     }
 
@@ -80,8 +82,30 @@ static void pmac_ide_atapi_transfer_cb(void *opaque, int ret)
         s->io_buffer_index &= 0x7ff;
     }
 
+    s->io_buffer_size = io->len;
+
+    MACIO_DPRINTF("remainder: %d io->len: %d size: %d\n", io->remainder_len,
+                  io->len, s->packet_transfer_size);
+    if (io->remainder_len && io->len) {
+        /* guest wants the rest of its previous transfer */
+        int remainder_len = MIN(io->remainder_len, io->len);
+
+        MACIO_DPRINTF("copying remainder %d bytes\n", remainder_len);
+
+        cpu_physical_memory_write(io->addr, io->remainder + 0x200 -
+                                  remainder_len, remainder_len);
+
+        io->addr += remainder_len;
+        io->len -= remainder_len;
+        s->io_buffer_size = remainder_len;
+        io->remainder_len -= remainder_len;
+        /* treat remainder as individual transfer, start again */
+        pmac_ide_atapi_transfer_cb(opaque, 0);
+        return;
+    }
+
     /* end of transfer ? */
-    if (s->packet_transfer_size <= 0) {
+    if (!s->packet_transfer_size) {
         MACIO_DPRINTF("end of transfer\n");
         ide_atapi_cmd_ok(s);
         m->dma_active = false;
@@ -95,14 +119,40 @@ static void pmac_ide_atapi_transfer_cb(void *opaque, int ret)
 
     /* launch next transfer */
 
-    MACIO_DPRINTF("io->len = %#x\n", io->len);
+    /* handle unaligned accesses first, get them over with and only do the
+       remaining bulk transfer using our async DMA helpers */
+    unaligned = io->len & 0x1ff;
+    if (unaligned) {
+        int sector_num = (s->lba << 2) + (s->io_buffer_index >> 9);
+        int nsector = io->len >> 9;
 
-    s->io_buffer_size = io->len;
+        MACIO_DPRINTF("precopying unaligned %d bytes to %#lx\n",
+                      unaligned, io->addr + io->len - unaligned);
+
+        bdrv_read(s->bs, sector_num + nsector, io->remainder, 1);
+        cpu_physical_memory_write(io->addr + io->len - unaligned,
+                                  io->remainder, unaligned);
+
+        io->len -= unaligned;
+    }
+
+    MACIO_DPRINTF("io->len = %#x\n", io->len);
 
     qemu_sglist_init(&s->sg, io->len / MACIO_PAGE_SIZE + 1,
                      &address_space_memory);
     qemu_sglist_add(&s->sg, io->addr, io->len);
-    io->addr += io->len;
+    io->addr += s->io_buffer_size;
+    io->remainder_len = MIN(s->packet_transfer_size - s->io_buffer_size,
+                            (0x200 - unaligned) & 0x1ff);
+    MACIO_DPRINTF("set remainder to: %d\n", io->remainder_len);
+
+    /* We would read no data from the block layer, thus not get a callback.
+       Just fake completion manually. */
+    if (!io->len) {
+        pmac_ide_atapi_transfer_cb(opaque, 0);
+        return;
+    }
+
     io->len = 0;
 
     MACIO_DPRINTF("sector_num=%d size=%d, cmd_cmd=%d\n",
@@ -125,14 +175,16 @@ static void pmac_ide_transfer_cb(void *opaque, int ret)
     DBDMA_io *io = opaque;
     MACIOIDEState *m = io->opaque;
     IDEState *s = idebus_active_if(&m->bus);
-    int n;
+    int n = 0;
     int64_t sector_num;
+    int unaligned;
 
     if (ret < 0) {
         MACIO_DPRINTF("DMA error\n");
         m->aiocb = NULL;
         qemu_sglist_destroy(&s->sg);
         ide_dma_error(s);
+        io->remainder_len = 0;
         goto done;
     }
 
@@ -155,8 +207,34 @@ static void pmac_ide_transfer_cb(void *opaque, int ret)
         s->nsector -= n;
     }
 
+    MACIO_DPRINTF("remainder: %d io->len: %d nsector: %d sector_num: %ld\n",
+                  io->remainder_len, io->len, s->nsector, sector_num);
+    if (io->remainder_len && io->len) {
+        /* guest wants the rest of its previous transfer */
+        int remainder_len = MIN(io->remainder_len, io->len);
+        uint8_t *p = &io->remainder[0x200 - remainder_len];
+
+        MACIO_DPRINTF("copying remainder %d bytes at %#lx\n",
+                      remainder_len, io->addr);
+
+        switch (s->dma_cmd) {
+        case IDE_DMA_READ:
+            cpu_physical_memory_write(io->addr, p, remainder_len);
+            break;
+        case IDE_DMA_WRITE:
+            cpu_physical_memory_read(io->addr, p, remainder_len);
+            bdrv_write(s->bs, sector_num - 1, io->remainder, 1);
+            break;
+        case IDE_DMA_TRIM:
+            break;
+        }
+        io->addr += remainder_len;
+        io->len -= remainder_len;
+        io->remainder_len -= remainder_len;
+    }
+
     /* end of transfer ? */
-    if (s->nsector == 0) {
+    if (s->nsector == 0 && !io->remainder_len) {
         MACIO_DPRINTF("end of transfer\n");
         s->status = READY_STAT | SEEK_STAT;
         ide_set_irq(s->bus);
@@ -174,13 +252,49 @@ static void pmac_ide_transfer_cb(void *opaque, int ret)
     s->io_buffer_index = 0;
     s->io_buffer_size = io->len;
 
+    /* handle unaligned accesses first, get them over with and only do the
+       remaining bulk transfer using our async DMA helpers */
+    unaligned = io->len & 0x1ff;
+    if (unaligned) {
+        int nsector = io->len >> 9;
+
+        MACIO_DPRINTF("precopying unaligned %d bytes to %#lx\n",
+                      unaligned, io->addr + io->len - unaligned);
+
+        switch (s->dma_cmd) {
+        case IDE_DMA_READ:
+            bdrv_read(s->bs, sector_num + nsector, io->remainder, 1);
+            cpu_physical_memory_write(io->addr + io->len - unaligned,
+                                      io->remainder, unaligned);
+            break;
+        case IDE_DMA_WRITE:
+            /* cache the contents in our io struct */
+            cpu_physical_memory_read(io->addr + io->len - unaligned,
+                                     io->remainder, unaligned);
+            break;
+        case IDE_DMA_TRIM:
+            break;
+        }
+
+        io->len -= unaligned;
+    }
 
     MACIO_DPRINTF("io->len = %#x\n", io->len);
 
     qemu_sglist_init(&s->sg, io->len / MACIO_PAGE_SIZE + 1,
                      &address_space_memory);
     qemu_sglist_add(&s->sg, io->addr, io->len);
-    io->addr += io->len;
+    io->addr += io->len + unaligned;
+    io->remainder_len = (0x200 - unaligned) & 0x1ff;
+    MACIO_DPRINTF("set remainder to: %d\n", io->remainder_len);
+
+    /* We would read no data from the block layer, thus not get a callback.
+       Just fake completion manually. */
+    if (!io->len) {
+        pmac_ide_transfer_cb(opaque, 0);
+        return;
+    }
+
     io->len = 0;
 
     MACIO_DPRINTF("sector_num=%" PRId64 " n=%d, nsector=%d, cmd_cmd=%d\n",
diff --git a/include/hw/ppc/mac_dbdma.h b/include/hw/ppc/mac_dbdma.h
index 8ad1b6e..ad4d826 100644
--- a/include/hw/ppc/mac_dbdma.h
+++ b/include/hw/ppc/mac_dbdma.h
@@ -39,6 +39,9 @@ struct DBDMA_io {
     DBDMA_end dma_end;
     /* DMA is in progress, don't start another one */
     int processing;
+    /* unaligned last sector of a request */
+    uint8_t remainder[0x200];
+    int remainder_len;
 };
 
 /*
-- 
1.8.1.4