diff mbox

[v17,01/10] LIB: Introduce a generic PIO mapping method

Message ID 20180403143909.GA21171@ulmo (mailing list archive)
State Not Applicable, archived
Headers show

Commit Message

Thierry Reding April 3, 2018, 2:39 p.m. UTC
On Tue, Apr 03, 2018 at 04:04:10PM +0200, Thierry Reding wrote:
> On Thu, Mar 15, 2018 at 02:15:50AM +0800, John Garry wrote:
> > From: Zhichang Yuan <yuanzhichang@hisilicon.com>
> > 
> > In commit 41f8bba7f555 ("of/pci: Add pci_register_io_range() and
> > pci_pio_to_address()"), a new I/O space management was supported. With
> > that driver, the I/O ranges configured for PCI/PCIe hosts on some
> > architectures can be mapped to logical PIO, converted easily between
> > CPU address and the corresponding logicial PIO. Based on this, PCI
> > I/O devices can be accessed in a memory read/write way through the
> > unified in/out accessors.
> > 
> > But on some archs/platforms, there are bus hosts which access I/O
> > peripherals with host-local I/O port addresses rather than memory
> > addresses after memory-mapped.
> > 
> > To support those devices, a more generic I/O mapping method is introduced
> > here. Through this patch, both the CPU addresses and the host-local port
> > can be mapped into the logical PIO space with different logical/fake PIOs.
> > After this, all the I/O accesses to either PCI MMIO devices or host-local
> > I/O peripherals can be unified into the existing I/O accessors defined in
> > asm-generic/io.h and be redirected to the right device-specific hooks
> > based on the input logical PIO.
> > 
> > Signed-off-by: Zhichang Yuan <yuanzhichang@hisilicon.com>
> > Signed-off-by: Gabriele Paoloni <gabriele.paoloni@huawei.com>
> > Signed-off-by: John Garry <john.garry@huawei.com>
> > Reviewed-by: Andy Shevchenko <andy.shevchenko@gmail.com>
> > Tested-by: dann frazier <dann.frazier@canonical.com>
> > ---
> >  include/asm-generic/io.h  |   2 +
> >  include/linux/logic_pio.h | 124 ++++++++++++++++++++
> >  lib/Kconfig               |  15 +++
> >  lib/Makefile              |   2 +
> >  lib/logic_pio.c           | 282 ++++++++++++++++++++++++++++++++++++++++++++++
> >  5 files changed, 425 insertions(+)
> >  create mode 100644 include/linux/logic_pio.h
> >  create mode 100644 lib/logic_pio.c
> > 
> [...]
> > diff --git a/lib/logic_pio.c b/lib/logic_pio.c
> > new file mode 100644
> > index 0000000..8394c2d
> > --- /dev/null
> > +++ b/lib/logic_pio.c
> > @@ -0,0 +1,282 @@
> > +// SPDX-License-Identifier: GPL-2.0+
> > +/*
> > + * Copyright (C) 2017 Hisilicon Limited, All Rights Reserved.
> > + * Author: Gabriele Paoloni <gabriele.paoloni@huawei.com>
> > + * Author: Zhichang Yuan <yuanzhichang@hisilicon.com>
> > + *
> > + */
> > +
> > +#define pr_fmt(fmt)	"LOGIC PIO: " fmt
> > +
> > +#include <linux/of.h>
> > +#include <linux/io.h>
> > +#include <linux/logic_pio.h>
> > +#include <linux/mm.h>
> > +#include <linux/rculist.h>
> > +#include <linux/sizes.h>
> > +#include <linux/slab.h>
> > +
> > +/* The unique hardware address list. */
> > +static LIST_HEAD(io_range_list);
> > +static DEFINE_MUTEX(io_range_mutex);
> > +
> > +/* Consider a kernel general helper for this */
> > +#define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
> > +
> > +/**
> > + * logic_pio_register_range - register logical PIO range for a host
> > + * @new_range: pointer to the io range to be registered.
> > + *
> > + * returns 0 on success, the error code in case of failure
> > + *
> > + * Register a new io range node in the io range list.
> > + */
> > +int logic_pio_register_range(struct logic_pio_hwaddr *new_range)
> > +{
> > +	struct logic_pio_hwaddr *range;
> > +	resource_size_t start = new_range->hw_start;
> > +	resource_size_t end = new_range->hw_start + new_range->size;
> > +	resource_size_t mmio_sz = 0;
> > +	resource_size_t iio_sz = MMIO_UPPER_LIMIT;
> > +	int ret = 0;
> > +
> > +	if (!new_range || !new_range->fwnode || !new_range->size)
> > +		return -EINVAL;
> > +
> > +	mutex_lock(&io_range_mutex);
> > +	list_for_each_entry_rcu(range, &io_range_list, list) {
> > +		if (range->fwnode == new_range->fwnode) {
> > +			/* range already there */
> > +			ret = -EFAULT;
> > +			goto end_register;
> > +		}
> 
> This is the -EFAULT that propagates to pci-tegra.c's ->probe() and fails
> to bind the driver.
> 
> I'm not exactly sure what's causing the duplicate here because it's
> rather difficult to get at something useful from just the ->fwnode, but
> I'm fairly sure that the reason this breaks is because the Tegra driver
> will defer probe due to some regulators that aren't available on the
> first try. Given the above code and the rest of this file, I can't see a
> way to "fix" the driver and remove the I/O range on failure.
> 
> This is doubly bad because this doesn't only leak the ranges on probe
> deferral, but also on driver unload, and we just added support for
> building the Tegra driver as a loadable module, so these are actually
> cases that can happen in regular uses of the driver.
> 
> I have no idea on how to fix this. Anyone know of a quick fix to restore
> PCI for Tegra other than reverting all of these changes?
> 
> I suppose an API could be added to unregister the range, but the calling
> sequence is rather obfuscated, so removing the range will look totally
> asymmetric, I'm afraid.
> 
> Here's the call stack:
> 
> 	tegra_pcie_probe()
> 	tegra_pcie_parse_dt()
> 	of_pci_range_to_resource()
> 	pci_register_io_range()
> 	logic_pio_register_range()
> 
> So the range here is registered as part of a resource parsing function,
> which is supposed to not have any side-effects. There's no equivalent of
> that parsing routine (i.e. no "unparse" function that would undo the
> effects of parsing).
> 
> Perhaps a cleaner way would be to decouple the parsing from the actual
> request step that has the side-effect.
> 
> Going back in history a little, it looks like even before this commit
> the I/O range registration was triggered by the parsing code and even
> the range leak was there, except that it caused pci_register_io_range()
> to return 0 rather than -EFAULT. Perhaps the quickest fix for this would
> be to do the same in the new code and restore drivers that accidentally
> depend on this behaviour.

I can confirm that the following fixes the issue for me, though I don't
think it's a very clean fix given that the range will remain requested
forever, even if the driver is gone. But since that's already been the
case for quite a while, probably something that can be fixed separately.

Cc'ing linux-tegra for visibility.

Thierry

--- >8 ---
diff --git a/lib/logic_pio.c b/lib/logic_pio.c
index 29cedeadb397..4664b87e1c5f 100644

Comments

John Garry April 3, 2018, 4:01 p.m. UTC | #1
>>> +int logic_pio_register_range(struct logic_pio_hwaddr *new_range)
>>> +{
>>> +	struct logic_pio_hwaddr *range;
>>> +	resource_size_t start = new_range->hw_start;
>>> +	resource_size_t end = new_range->hw_start + new_range->size;
>>> +	resource_size_t mmio_sz = 0;
>>> +	resource_size_t iio_sz = MMIO_UPPER_LIMIT;
>>> +	int ret = 0;
>>> +
>>> +	if (!new_range || !new_range->fwnode || !new_range->size)
>>> +		return -EINVAL;
>>> +
>>> +	mutex_lock(&io_range_mutex);
>>> +	list_for_each_entry_rcu(range, &io_range_list, list) {
>>> +		if (range->fwnode == new_range->fwnode) {
>>> +			/* range already there */
>>> +			ret = -EFAULT;
>>> +			goto end_register;
>>> +		}
>>

Hi Thierry,

>> This is the -EFAULT that propagates to pci-tegra.c's ->probe() and fails
>> to bind the driver.
>>
>> I'm not exactly sure what's causing the duplicate here because it's
>> rather difficult to get at something useful from just the ->fwnode, but
>> I'm fairly sure that the reason this breaks is because the Tegra driver
>> will defer probe due to some regulators that aren't available on the
>> first try. Given the above code and the rest of this file, I can't see a
>> way to "fix" the driver and remove the I/O range on failure.
>>
>> This is doubly bad because this doesn't only leak the ranges on probe
>> deferral, but also on driver unload, and we just added support for
>> building the Tegra driver as a loadable module, so these are actually
>> cases that can happen in regular uses of the driver.
>>
>> I have no idea on how to fix this. Anyone know of a quick fix to restore
>> PCI for Tegra other than reverting all of these changes?
>>
>> I suppose an API could be added to unregister the range, but the calling
>> sequence is rather obfuscated, so removing the range will look totally
>> asymmetric, I'm afraid.
>>
>> Here's the call stack:
>>
>> 	tegra_pcie_probe()
>> 	tegra_pcie_parse_dt()
>> 	of_pci_range_to_resource()
>> 	pci_register_io_range()
>> 	logic_pio_register_range()
>>
>> So the range here is registered as part of a resource parsing function,
>> which is supposed to not have any side-effects. There's no equivalent of
>> that parsing routine (i.e. no "unparse" function that would undo the
>> effects of parsing).
>>
>> Perhaps a cleaner way would be to decouple the parsing from the actual
>> request step that has the side-effect.

This could be added if we agreed that it would be useful.

>>
>> Going back in history a little, it looks like even before this commit
>> the I/O range registration was triggered by the parsing code and even
>> the range leak was there, except that it caused pci_register_io_range()
>> to return 0 rather than -EFAULT. Perhaps the quickest fix for this would
>> be to do the same in the new code and restore drivers that accidentally
>> depend on this behaviour.
>
> I can confirm that the following fixes the issue for me, though I don't
> think it's a very clean fix given that the range will remain requested
> forever, even if the driver is gone. But since that's already been the
> case for quite a while, probably something that can be fixed separately.
>

Right, there was no way to deregister the range previously.  From 
looking at the history here I see no reason to not support it.

As for this patch, as you said, the only difference is that we fault on 
trying to register the same range again. So this solution seems reasonable.

On another point, for the tegra driver, is it possible to defer earlier 
in the probe, before these currently irreversible steps are taken?

Thanks very much,
John

> Cc'ing linux-tegra for visibility.
>
> Thierry
>
> --- >8 ---
> diff --git a/lib/logic_pio.c b/lib/logic_pio.c
> index 29cedeadb397..4664b87e1c5f 100644
> --- a/lib/logic_pio.c
> +++ b/lib/logic_pio.c
> @@ -46,7 +46,6 @@ int logic_pio_register_range(struct logic_pio_hwaddr *new_range)
>  	list_for_each_entry_rcu(range, &io_range_list, list) {
>  		if (range->fwnode == new_range->fwnode) {
>  			/* range already there */
> -			ret = -EFAULT;
>  			goto end_register;
>  		}
>  		if (range->flags == LOGIC_PIO_CPU_MMIO &&
>


--
To unsubscribe from this list: send the line "unsubscribe linux-acpi" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Thierry Reding April 3, 2018, 4:37 p.m. UTC | #2
On Tue, Apr 03, 2018 at 05:01:37PM +0100, John Garry wrote:
> > > > +int logic_pio_register_range(struct logic_pio_hwaddr *new_range)
> > > > +{
> > > > +	struct logic_pio_hwaddr *range;
> > > > +	resource_size_t start = new_range->hw_start;
> > > > +	resource_size_t end = new_range->hw_start + new_range->size;
> > > > +	resource_size_t mmio_sz = 0;
> > > > +	resource_size_t iio_sz = MMIO_UPPER_LIMIT;
> > > > +	int ret = 0;
> > > > +
> > > > +	if (!new_range || !new_range->fwnode || !new_range->size)
> > > > +		return -EINVAL;
> > > > +
> > > > +	mutex_lock(&io_range_mutex);
> > > > +	list_for_each_entry_rcu(range, &io_range_list, list) {
> > > > +		if (range->fwnode == new_range->fwnode) {
> > > > +			/* range already there */
> > > > +			ret = -EFAULT;
> > > > +			goto end_register;
> > > > +		}
> > > 
> 
> Hi Thierry,
> 
> > > This is the -EFAULT that propagates to pci-tegra.c's ->probe() and fails
> > > to bind the driver.
> > > 
> > > I'm not exactly sure what's causing the duplicate here because it's
> > > rather difficult to get at something useful from just the ->fwnode, but
> > > I'm fairly sure that the reason this breaks is because the Tegra driver
> > > will defer probe due to some regulators that aren't available on the
> > > first try. Given the above code and the rest of this file, I can't see a
> > > way to "fix" the driver and remove the I/O range on failure.
> > > 
> > > This is doubly bad because this doesn't only leak the ranges on probe
> > > deferral, but also on driver unload, and we just added support for
> > > building the Tegra driver as a loadable module, so these are actually
> > > cases that can happen in regular uses of the driver.
> > > 
> > > I have no idea on how to fix this. Anyone know of a quick fix to restore
> > > PCI for Tegra other than reverting all of these changes?
> > > 
> > > I suppose an API could be added to unregister the range, but the calling
> > > sequence is rather obfuscated, so removing the range will look totally
> > > asymmetric, I'm afraid.
> > > 
> > > Here's the call stack:
> > > 
> > > 	tegra_pcie_probe()
> > > 	tegra_pcie_parse_dt()
> > > 	of_pci_range_to_resource()
> > > 	pci_register_io_range()
> > > 	logic_pio_register_range()
> > > 
> > > So the range here is registered as part of a resource parsing function,
> > > which is supposed to not have any side-effects. There's no equivalent of
> > > that parsing routine (i.e. no "unparse" function that would undo the
> > > effects of parsing).
> > > 
> > > Perhaps a cleaner way would be to decouple the parsing from the actual
> > > request step that has the side-effect.
> 
> This could be added if we agreed that it would be useful.

I guess in most cases these ranges will be static at least during one
boot. But it still feels like this should be removed when the driver
goes away. While this may not depend on data by the driver, and hence
won't cause a crash or anything, it just seems wrong to leave it
around when the driver no longer isn't.

> > > Going back in history a little, it looks like even before this commit
> > > the I/O range registration was triggered by the parsing code and even
> > > the range leak was there, except that it caused pci_register_io_range()
> > > to return 0 rather than -EFAULT. Perhaps the quickest fix for this would
> > > be to do the same in the new code and restore drivers that accidentally
> > > depend on this behaviour.
> > 
> > I can confirm that the following fixes the issue for me, though I don't
> > think it's a very clean fix given that the range will remain requested
> > forever, even if the driver is gone. But since that's already been the
> > case for quite a while, probably something that can be fixed separately.
> > 
> 
> Right, there was no way to deregister the range previously.  From looking at
> the history here I see no reason to not support it.
> 
> As for this patch, as you said, the only difference is that we fault on
> trying to register the same range again. So this solution seems reasonable.

Okay, I can turn this into a proper patch to fix this up. I suspect that
other drivers may be subject to the same regression. For the longer term
I think it'd be better to properly undo the registration on failure and
removal, but I suspect that it'd be quite a bit of work and not suitable
for v4.17 anymore.

> On another point, for the tegra driver, is it possible to defer earlier in
> the probe, before these currently irreversible steps are taken?

I'm sure it'd be possible. But it would be quite involved, I think. The
reason the code is the way it is is because parsing the DT didn't use to
have side-effects.

Also, I don't think it would buy us much because the probe can still
defer (or at least fail) as late as pci_scan_root_bus_bridge(). Even if
we work around the probe deferral by moving the DT parsing to a later
point we could easily run into a situation where the entry remains in
place and a subsequent attempt to reload the driver would then fail in
the same way as if we were deferring probe.

Thierry
John Garry April 3, 2018, 5:02 p.m. UTC | #3
On 03/04/2018 17:37, Thierry Reding wrote:
> On Tue, Apr 03, 2018 at 05:01:37PM +0100, John Garry wrote:
>>>>> +int logic_pio_register_range(struct logic_pio_hwaddr *new_range)
>>>>> +{
>>>>> +	struct logic_pio_hwaddr *range;
>>>>> +	resource_size_t start = new_range->hw_start;
>>>>> +	resource_size_t end = new_range->hw_start + new_range->size;
>>>>> +	resource_size_t mmio_sz = 0;
>>>>> +	resource_size_t iio_sz = MMIO_UPPER_LIMIT;
>>>>> +	int ret = 0;
>>>>> +
>>>>> +	if (!new_range || !new_range->fwnode || !new_range->size)
>>>>> +		return -EINVAL;
>>>>> +
>>>>> +	mutex_lock(&io_range_mutex);
>>>>> +	list_for_each_entry_rcu(range, &io_range_list, list) {
>>>>> +		if (range->fwnode == new_range->fwnode) {
>>>>> +			/* range already there */
>>>>> +			ret = -EFAULT;
>>>>> +			goto end_register;
>>>>> +		}
>>>>
>>
>> Hi Thierry,
>>
>>>> This is the -EFAULT that propagates to pci-tegra.c's ->probe() and fails
>>>> to bind the driver.
>>>>
>>>> I'm not exactly sure what's causing the duplicate here because it's
>>>> rather difficult to get at something useful from just the ->fwnode, but
>>>> I'm fairly sure that the reason this breaks is because the Tegra driver
>>>> will defer probe due to some regulators that aren't available on the
>>>> first try. Given the above code and the rest of this file, I can't see a
>>>> way to "fix" the driver and remove the I/O range on failure.
>>>>
>>>> This is doubly bad because this doesn't only leak the ranges on probe
>>>> deferral, but also on driver unload, and we just added support for
>>>> building the Tegra driver as a loadable module, so these are actually
>>>> cases that can happen in regular uses of the driver.
>>>>
>>>> I have no idea on how to fix this. Anyone know of a quick fix to restore
>>>> PCI for Tegra other than reverting all of these changes?
>>>>
>>>> I suppose an API could be added to unregister the range, but the calling
>>>> sequence is rather obfuscated, so removing the range will look totally
>>>> asymmetric, I'm afraid.
>>>>
>>>> Here's the call stack:
>>>>
>>>> 	tegra_pcie_probe()
>>>> 	tegra_pcie_parse_dt()
>>>> 	of_pci_range_to_resource()
>>>> 	pci_register_io_range()
>>>> 	logic_pio_register_range()
>>>>
>>>> So the range here is registered as part of a resource parsing function,
>>>> which is supposed to not have any side-effects. There's no equivalent of
>>>> that parsing routine (i.e. no "unparse" function that would undo the
>>>> effects of parsing).
>>>>
>>>> Perhaps a cleaner way would be to decouple the parsing from the actual
>>>> request step that has the side-effect.
>>
>> This could be added if we agreed that it would be useful.
>
> I guess in most cases these ranges will be static at least during one
> boot. But it still feels like this should be removed when the driver
> goes away. While this may not depend on data by the driver, and hence
> won't cause a crash or anything, it just seems wrong to leave it
> around when the driver no longer isn't.

That sounds reasonable, considering we do unmap the iospace when we 
release - so it looks like currently we're leaving some IO range 
reserved which does not have a mapping.

However this change seems non-trivial, considering we're now even 
coupling the PIO range registration into DT parsing.

>
>>>> Going back in history a little, it looks like even before this commit
>>>> the I/O range registration was triggered by the parsing code and even
>>>> the range leak was there, except that it caused pci_register_io_range()
>>>> to return 0 rather than -EFAULT. Perhaps the quickest fix for this would
>>>> be to do the same in the new code and restore drivers that accidentally
>>>> depend on this behaviour.
>>>
>>> I can confirm that the following fixes the issue for me, though I don't
>>> think it's a very clean fix given that the range will remain requested
>>> forever, even if the driver is gone. But since that's already been the
>>> case for quite a while, probably something that can be fixed separately.
>>>
>>
>> Right, there was no way to deregister the range previously.  From looking at
>> the history here I see no reason to not support it.
>>
>> As for this patch, as you said, the only difference is that we fault on
>> trying to register the same range again. So this solution seems reasonable.
>
> Okay, I can turn this into a proper patch to fix this up. I suspect that
> other drivers may be subject to the same regression. For the longer term
> I think it'd be better to properly undo the registration on failure and
> removal, but I suspect that it'd be quite a bit of work and not suitable
> for v4.17 anymore.

Thanks, I had started to put the patch together but if you're happy to 
continue then that's fine. Please let me know.

>
>> On another point, for the tegra driver, is it possible to defer earlier in
>> the probe, before these currently irreversible steps are taken?
>
> I'm sure it'd be possible. But it would be quite involved, I think. The
> reason the code is the way it is is because parsing the DT didn't use to
> have side-effects.
>
> Also, I don't think it would buy us much because the probe can still
> defer (or at least fail) as late as pci_scan_root_bus_bridge(). Even if
> we work around the probe deferral by moving the DT parsing to a later
> point we could easily run into a situation where the entry remains in
> place and a subsequent attempt to reload the driver would then fail in
> the same way as if we were deferring probe.
>
> Thierry
>

Thanks,
John

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

Patch

--- a/lib/logic_pio.c
+++ b/lib/logic_pio.c
@@ -46,7 +46,6 @@  int logic_pio_register_range(struct logic_pio_hwaddr *new_range)
 	list_for_each_entry_rcu(range, &io_range_list, list) {
 		if (range->fwnode == new_range->fwnode) {
 			/* range already there */
-			ret = -EFAULT;
 			goto end_register;
 		}
 		if (range->flags == LOGIC_PIO_CPU_MMIO &&