Skip to contentSkip to conversation

ide_dma_cancel will result in partial DMA transfer (resend #3)

PATCH In review Andrea Arcangeli 2 participants 3 messages Feb 18, 2009 – Feb 26, 2009

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 ide_dma_cancel will result in partial DMA transfer (resend #3)

From: Andrea Arcangeli <[email protected]>

The reason for not actually canceling the I/O is because with virtualization and lots of VM running, a guest fs may mistake a overload of the host, as an IDE timeout. So rather than canceling the I/O, it's safer to wait I/O completion and simulate that the I/O has completed just before the io cancellation was requested by the guest. This way if ntfs or an app writes data without checking for -EIO retval, and it thinks the write has succeeded, it's less likely to run into troubles. Similar issues for reads.

Furthermore because the DMA operation is splitted into many synchronous aio_read/write if there's more than one entry in the SG table, without this patch the DMA would be cancelled in the middle, something we've no idea if it happens on real hardware too or not. Overall this seems a great risk for zero gain.

This approach is sure safer than previous code given we can't pretend all guest fs code out there to check for errors and reply the DMA if it was completed partially, given a timeout would never materialize on a real harddisk unless there are defective blocks (and defective blocks are practically only an issue for reads never for writes in any recent hardware as writing to blocks is the way to fix them) or the harddisk breaks as a whole.

---

This is a resubmit of an old patch in my queue. Wonder if it'll ever be merged. I think it's obviously safer (especially once we've preadv/pwritev driven I/O) even if a noop.

ide.c +22 -2
Index: hw/ide.c
===================================================================
@@ -2878,8 +2878,28 @@
     printf("%s: 0x%08x\n", __func__, val);
 #endif
     if (!(val & BM_CMD_START)) {
-        /* XXX: do it better */
-        ide_dma_cancel(bm);
+        /*
+	 * We can't cancel Scatter Gather DMA in the middle of the
+	 * operation or a partial (not full) DMA transfer would reach
+	 * the storage so we wait for completion instead (we beahve
+	 * like if the DMA was complated by the time the guest trying
+	 * to cancel dma with bmdma_cmd_writeb with BM_CMD_START not
+	 * set).
+	 *
+	 * In the future we'll be able to safely cancel the I/O if the
+	 * whole DMA operation will be submitted to disk with a single
+	 * aio operation in the form of aio_readv/aio_writev
+	 * (supported by linux kernel AIO but not by glibc pthread aio
+	 * lib).
+	 */
+	if (bm->aiocb) {
+		QEMU_WARN("qemu_aio_flush called");
+		qemu_aio_flush();
+		if (bm->aiocb)
+			QEMU_WARN("aiocb still pending");
+		if (bm->status & BM_STATUS_DMAING)
+			QEMU_WARN("BM_STATUS_DMAING still pending");
+	}
         bm->cmd = val & 0x09;
     } else {
         if (!(bm->status & BM_STATUS_DMAING)) {
#1 lore raw

ide.c line 2896

Quoted 3 lines · Andrea Arcangeli
ide.c +2 -0
+	if (bm->aiocb) {
+		QEMU_WARN("qemu_aio_flush called");

This doesn't build (there is no QEMU_WARN). I think I've mentioned this before?

Regards,

Anthony Liguori

Quoted 11 lines
ide.c +6 -0
+		qemu_aio_flush();
+		if (bm->aiocb)
+			QEMU_WARN("aiocb still pending");
+		if (bm->status & BM_STATUS_DMAING)
+			QEMU_WARN("BM_STATUS_DMAING still pending");
+	}
         bm->cmd = val & 0x09;
     } else {
         if (!(bm->status & BM_STATUS_DMAING)) {
#2 lore raw
Quoted 7 lines · Anthony Liguori

Review

Quoted 3 lines · Andrea Arcangeli
(quoted diff) +2 -0
+	if (bm->aiocb) {
+		QEMU_WARN("qemu_aio_flush called");

This doesn't build (there is no QEMU_WARN). I think I've mentioned this before?

First patch that triggered some replies didn't have this problem. Must have been a mistake when resending #2 that didn't generate replies. This one below should build too.

------
Subject: avoid cancelling ide dma
From: Andrea Arcangeli <[email protected]>

The reason for not actually canceling the I/O is because with virtualization and lots of VM running, a guest fs may mistake a overload of the host, as an IDE timeout. So rather than canceling the I/O, it's safer to wait I/O completion and simulate that the I/O has completed just before the io cancellation was requested by the guest. This way if ntfs or an app writes data without checking for -EIO retval, and it thinks the write has succeeded, it's less likely to run into troubles. Similar issues for reads.

Furthermore because the DMA operation is splitted into many synchronous aio_read/write if there's more than one entry in the SG table, without this patch the DMA would be cancelled in the middle, something we've no idea if it happens on real hardware too or not. Overall this seems a great risk for zero gain.

This approach is sure safer than previous code given we can't pretend all guest fs code out there to check for errors and reply the DMA if it was completed partially, given a timeout would never materialize on a real harddisk unless there are defective blocks (and defective blocks are practically only an issue for reads never for writes in any recent hardware as writing to blocks is the way to fix them) or the harddisk breaks as a whole.

---

ide.c +21 -2
Index: hw/ide.c
===================================================================
@@ -2983,8 +2983,27 @@
     printf("%s: 0x%08x\n", __func__, val);
 #endif
     if (!(val & BM_CMD_START)) {
-        /* XXX: do it better */
-        ide_dma_cancel(bm);
+        /*
+	 * We can't cancel Scatter Gather DMA in the middle of the
+	 * operation or a partial (not full) DMA transfer would reach
+	 * the storage so we wait for completion instead (we beahve
+	 * like if the DMA was complated by the time the guest trying
+	 * to cancel dma with bmdma_cmd_writeb with BM_CMD_START not
+	 * set).
+	 *
+	 * In the future we'll be able to safely cancel the I/O if the
+	 * whole DMA operation will be submitted to disk with a single
+	 * aio operation with preadv/pwritev.
+	 */
+	if (bm->aiocb) {
+		qemu_aio_flush();
+#ifdef DEBUG_IDE
+		if (bm->aiocb)
+			printf("aiocb still pending");
+		if (bm->status & BM_STATUS_DMAING)
+			printf("BM_STATUS_DMAING still pending");
+#endif
+	}
         bm->cmd = val & 0x09;
     } else {
         if (!(bm->status & BM_STATUS_DMAING)) {