On 19.9.2015 01:26, Matthew R. Ochs wrote:

quoted

On Sep 18, 2015, at 6:59 AM, Tomas Henzl [off-list ref] wrote:
On 17.9.2015 19:16, Matthew R. Ochs wrote:

quoted

quoted

On Sep 17, 2015, at 7:38 AM, Tomas Henzl [off-list ref] wrote:

On 16.9.2015 18:53, Matthew R. Ochs wrote:

quoted

Interrupt processing can run in parallel to a remove operation. This
can lead to a condition where the interrupt handler is processing with
memory that has been freed.

To avoid processing an interrupt while memory may be yanked, check for
removal while in the interrupt handler. Bail when removal is imminent.

Signed-off-by: Matthew R. Ochs <redacted>
Signed-off-by: Manoj N. Kumar <redacted>
---
drivers/scsi/cxlflash/common.h |  2 ++
drivers/scsi/cxlflash/main.c   | 21 +++++++++++++++------
2 files changed, 17 insertions(+), 6 deletions(-)

diff --git a/drivers/scsi/cxlflash/common.h b/drivers/scsi/cxlflash/common.h
index 1abe4e0..03d2cc6 100644
--- a/drivers/scsi/cxlflash/common.h
+++ b/drivers/scsi/cxlflash/common.h

@@ -103,6 +103,8 @@ struct cxlflash_cfg {

	enum cxlflash_lr_state lr_state;
	int lr_port;

+	atomic_t remove_active;
+
	struct cxl_afu *cxl_afu;

	struct pci_pool *cxlflash_cmd_pool;

diff --git a/drivers/scsi/cxlflash/main.c b/drivers/scsi/cxlflash/main.c
index 6e85c77..89ee648 100644
--- a/drivers/scsi/cxlflash/main.c
+++ b/drivers/scsi/cxlflash/main.c

@@ -892,6 +892,7 @@ static void cxlflash_remove(struct pci_dev *pdev)

	spin_unlock_irqrestore(&cfg->tmf_waitq.lock, lock_flags);

	cfg->state = STATE_FAILTERM;
+	atomic_inc(&cfg->remove_active);

Hi Matthew,
you could just call term_afu at this point, this way you don't
need an additional check in all irq functions.
Cheers,
Tomas

Hi Tomas,

We actually do call term_afu() a few lines down from here. I don't follow
how moving it here would help things.

When you disable ints sooner (that is what term_afu does ?) you'll get no
more ints later isn't this what you want?

Correct, that's what we want.

quoted

quoted

The reason for the atomic was to provide something lightweight that we
could check _inside_ the processing loop for the read-response queue
handler. A check outside that loop doesn't really provide much in terms
of closing or narrowing down the window of when freed memory can be
accessed.

As David Laight correctly pointed out, this approach does not completely
close the window. We'd need something heavier to fully protect (e.g. a lock
to wrap around the entire loop). I will look into adding this in a future cycle
when I can adequately test.

term_afu calls free_irq and this function
does not return until any executing interrupts for have completed.
This is the sync mechanism you need, it's lightweight
(does not add an additional check to your irq functions)
and closes the race window completely.

Thanks for clarifying!

I looked at this closer and you are correct, free_irq() guarantees not
to return until the interrupt handler has completed. The current location
of term_afu() is appropriate as the memory that the handler touches is
not freed until the very end [by free_mem() and scsi_host_put()] of the
remove. Thus we can simply ignore this patch (I'll remove it in a v3).

OK. In some future patch please reorganize the remove function so,
that it follows the template described in Documentation/PCI/pci.txt :
	Disable the device from generating IRQs
	Release the IRQ (free_irq())
	Stop all DMA activity
	Release DMA buffers (both streaming and coherent)
	Unregister from other subsystems (e.g. scsi or netdev)
	Release MMIO/IOP resources
	Disable the device

-matt


--
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html