Skip to contentSkip to conversation

MMC: Add 400ms to CAFE SD controller resume path

PATCH In review Deepak Saxena 3 participants 6 messages May 13, 2009 – Jun 13, 2009

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 MMC: Add 400ms to CAFE SD controller resume path

The CAFE SD controller takes a long time to completely resume and w/o this patch, we do not redetect an existing card but instead detect it as a new one. Even with MMC_UNSAFE_RESUME enabled, this leads to the partition table on the device being wiped out.

See http://dev.laptop.org/ticket/6532 for gory details.

---

drivers/mmc/host/sdhci-pci.c +16 -0
index 65be279..9941bb0 100644
@@ -105,10 +105,26 @@ static const struct sdhci_pci_fixes sdhci_ene_714 = {
 			  SDHCI_QUIRK_BROKEN_DMA,
 };
 
+
+/*
+ * This 400ms timeout is needed to ensure that the chip has proper time to
+ * wake up an redetect an already inserted card. Without this, we end up
+ * corrupting the partition table on the device.
+ *
+ * See http://dev.laptop.org/ticket/6532 for lots of ugly details.
+ */
+static int cafe_resume(struct sdhci_pci_chip *chip)
+{
+	msleep(400);
+
+	return 0;
+}
+
 static const struct sdhci_pci_fixes sdhci_cafe = {
 	.quirks		= SDHCI_QUIRK_NO_SIMULT_VDD_AND_POWER |
 			  SDHCI_QUIRK_NO_BUSY_IRQ |
 			  SDHCI_QUIRK_BROKEN_TIMEOUT_VAL,
+	.resume		= cafe_resume,
 };
 
 static int jmicron_pmos(struct sdhci_pci_chip *chip, int on)
Signature · 2 lines
In the end, they will not say, "those were dark times,"  they will ask
"why were their poets silent?" - Bertold Brecht
#1 lore raw
Quoted 11 lines · On Wed, 13 May 2009 20:05:56 +0000 Deepak Saxena <[email protected]> wrote:

The CAFE SD controller takes a long time to completely resume and w/o this patch, we do not redetect an existing card but instead detect it as a new one. Even with MMC_UNSAFE_RESUME enabled, this leads to the partition table on the device being wiped out.

See http://dev.laptop.org/ticket/6532 for gory details.

Reading through that report, I don't believe you properly worked around the bug. You only avoid bug 1339, but that's only mildly related.

What this workaround does is to make sure that MMC_UNSAFE_RESUME actually works. But if you change cards during suspend, the VFS bug should reappear and you'll corrupt the partition table.

And as for 1339, has there been any more work done on why this problem doesn't appear in OFW? I'd be a lot happier if we could make things work without artificial delays.

RGds

Signature · 6 lines
     -- Pierre Ossman

  WARNING: This correspondence is being monitored by the
  Swedish government. Make sure your server uses encryption
  for SMTP traffic and consider using PGP for end-to-end
  encryption.
#2 lore raw
Quoted 11 lines · On Fri, 22 May 2009 13:51:50 +0200 Pierre Ossman <[email protected]> wrote:
Quoted 8 lines · On Wed, 13 May 2009 20:05:56 +0000 Deepak Saxena <[email protected]> wrote:

The CAFE SD controller takes a long time to completely resume and w/o this patch, we do not redetect an existing card but instead detect it as a new one. Even with MMC_UNSAFE_RESUME enabled, this leads to the partition table on the device being wiped out.

See http://dev.laptop.org/ticket/6532 for gory details.

hm, this sounds like a fairly serious problem, but the patch is nearly a year old.

Quoted 9 lines
Quoted 2 lines

Reading through that report, I don't believe you properly worked around the bug. You only avoid bug 1339, but that's only mildly related.

What this workaround does is to make sure that MMC_UNSAFE_RESUME actually works. But if you change cards during suspend, the VFS bug should reappear and you'll corrupt the partition table.

What do you think the VFS did wrong here?

Quoted 3 lines

And as for 1339, has there been any more work done on why this problem doesn't appear in OFW? I'd be a lot happier if we could make things work without artificial delays.

#3 lore raw
Quoted 12 lines · On Thu, 28 May 2009 01:36:29 -0700 Andrew Morton <[email protected]>…
Quoted 7 lines · On Fri, 22 May 2009 13:51:50 +0200 Pierre Ossman <[email protected]> wrote:

Reading through that report, I don't believe you properly worked around the bug. You only avoid bug 1339, but that's only mildly related.

What this workaround does is to make sure that MMC_UNSAFE_RESUME actually works. But if you change cards during suspend, the VFS bug should reappear and you'll corrupt the partition table.

What do you think the VFS did wrong here?

Might be the block layer as well. Somehow requests associated with an old block device end up on the queue of a new block device. I don't see how this can happen given Linux' device model, but somehow it does.

Rgds

Signature · 6 lines
     -- Pierre Ossman

  WARNING: This correspondence is being monitored by the
  Swedish government. Make sure your server uses encryption
  for SMTP traffic and consider using PGP for end-to-end
  encryption.
#4 lore raw

On May 28 2009, at 10:48, Pierre Ossman was caught saying:

Quoted 13 lines
Quoted 10 lines · On Thu, 28 May 2009 01:36:29 -0700 Andrew Morton <[email protected]>…
Quoted 7 lines · On Fri, 22 May 2009 13:51:50 +0200 Pierre Ossman <[email protected]> wrote:

Reading through that report, I don't believe you properly worked around the bug. You only avoid bug 1339, but that's only mildly related.

What this workaround does is to make sure that MMC_UNSAFE_RESUME actually works. But if you change cards during suspend, the VFS bug should reappear and you'll corrupt the partition table.

Quoted 6 lines
Quoted 2 lines

What do you think the VFS did wrong here?

Might be the block layer as well. Somehow requests associated with an old block device end up on the queue of a new block device. I don't see how this can happen given Linux' device model, but somehow it does.

The request is not ending up in the queue of the new device. If I recall correctyl What is happening is that userspace does an unmount on the old device when it receives the notification that it is gone (since it is not redected and a new entry is created). By the time that unmount is called, the various data structures for the device has been zeroed out but the various operation pointers are !NULL, so when we go back to write the superblock for the mounted partition, we overwrite the device's partition table as the partition offset is now 0. The main issue here is that the kernel is allowing an I/O to a device that is technically still there but not from the kernel's POV.

~Deepak

Signature · 2 lines
In the end, they will not say, "those were dark times,"  they will ask
"why were their poets silent?" - Bertold Brecht
#5 lore raw
Quoted 11 lines · On Thu, 4 Jun 2009 23:33:49 +0000 Deepak Saxena <[email protected]> wrote:

The request is not ending up in the queue of the new device. If I recall correctyl What is happening is that userspace does an unmount on the old device when it receives the notification that it is gone (since it is not redected and a new entry is created). By the time that unmount is called, the various data structures for the device has been zeroed out but the various operation pointers are !NULL, so when we go back to write the superblock for the mounted partition, we overwrite the device's partition table as the partition offset is now 0. The main issue here is that the kernel is allowing an I/O to a device that is technically still there but not from the kernel's POV.

But if the device is not there from the kernel's POV then it should be impossible to write to it. It seems this is a kernel bug though. See this thread:

http://marc.info/?t=124413860500006&r=1&w=2

Rgds

Signature · 6 lines
     -- Pierre Ossman

  WARNING: This correspondence is being monitored by the
  Swedish government. Make sure your server uses encryption
  for SMTP traffic and consider using PGP for end-to-end
  encryption.