This small patch-set, together with amd_iommu_v2 patch at
http://lists.linuxfoundation.org/pipermail/iommu/2014-December/011435.html
was created to solve the bug described at
https://bugzilla.kernel.org/show_bug.cgi?id=89661 (Kernel panic when
trying use amdkfd driver on Kaveri).
That bug appears only when radeon, amdkfd and amd_iommu_v2 are compiled
inside the kernel (not as modules). In that case, the correct loading
order, as determined by the exported symbol used by each driver, is
not enforced anymore and the kernel loads them based on who was linked
first. That makes radeon load first, amdkfd second and amd_iommu_v2
third.
Because the initialization of a device in amdkfd is initiated by radeon,
and can only be completed if amdkfd and amd_iommu_v2 were loaded and
initialized, then in the case mentioned above, this initalization fails
and there is a kernel panic as some pointers are not initialized but
used nontheless.
To solve this problem, amdkfd now checks if both it and amd_iommu_v2
were loaded before trying to initalize the device. If not, it enqueue
the work using a workqueue, which allows radeon to continue its device
initialization (because radeon calls amdkfd to initalize the device).
The work function schedules itself as long as amdkfd and amd_iommu_v2
were not initialized.
Detection of when the modules finished their initialization is done by
a simple variable that is initialized to 1 when the module_init function
is completed successfully. Other methods for detection were checked,
e.g. module_is_live() and MODULE_SOFTDEP(), but they were proved not
to work when all modules are compiled in the kernel image (which is
the problematic scenario to begin with).
Oded
Oded Gabbay (3):
amdkfd: Don't clear *kfd2kgd on kfd_module_init
amdkfd: Track when amdkfd init is complete
amdkfd: Use workqueue for GPU init
drivers/gpu/drm/amd/amdkfd/kfd_device.c | 72 +++++++++++++++++++++++++++++++--
drivers/gpu/drm/amd/amdkfd/kfd_module.c | 14 +++++--
drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 4 ++
3 files changed, 83 insertions(+), 7 deletions(-)
--
2.1.0
This patch adds a new function to amdkfd, which returns 1 if the amdkfd
initialization function has completed, and 0 otherwise.
This is necessary for the case when amdkfd and radeon are both compiled inside
the kernel image (not as modules). In that case, radeon probes the existing GPU
before amdkfd is even loaded. When radeon encounters an AMD GPU, it will pass
that information to amdkfd. However, if amdkfd is not loaded than that call
will fail and when amdkfd is eventually loaded, it won't know about that GPU.
Even if that call is delayed to a later stage, we will still need to know
whether amdkfd has already been initialized.
Note that when the two modules are compiled as modules, this situation can't
occur as the kernel enforces the order of loading.
Signed-off-by: Oded Gabbay <[email protected]>
---
drivers/gpu/drm/amd/amdkfd/kfd_module.c | 9 +++++++++
drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 2 ++
2 files changed, 11 insertions(+)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_module.c b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
index 236562f..848ebed 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_module.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
@@ -60,6 +60,13 @@ module_param(max_num_of_queues_per_process, int, 0444);
MODULE_PARM_DESC(max_num_of_queues_per_process,
"Kernel cmdline parameter that defines the amdkfd maximum number of supported queues per process");
+static int amdkfd_init_completed;
+
+int amdkfd_is_init_completed(void)
+{
+ return amdkfd_init_completed;
+}
+
bool kgd2kfd_init(unsigned interface_version,
const struct kfd2kgd_calls *f2g,
const struct kgd2kfd_calls **g2f)
@@ -128,6 +135,8 @@ static int __init kfd_module_init(void)
dev_info(kfd_device, "Initialized module\n");
+ amdkfd_init_completed = 1;
+
return 0;
err_topology:
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
index f9fb81e..01df7e6 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
@@ -153,6 +153,8 @@ struct kfd_dev {
bool interrupts_active;
};
+int amdkfd_is_init_completed(void);
+
/* KGD2KFD callbacks */
void kgd2kfd_exit(void);
struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd, struct pci_dev *pdev);
--
2.1.0
When amd_iommu_v2, amdkfd and radeon are all compiled inside the kernel image
(not as modules), radeon probes the existing GPU before amdkfd and amd_iommu_v2
are even loaded. When radeon encounters an AMD GPU, it will pass that
information to amdkfd. However, that call will fail and will cause a kernel BUG.
We could poll in radeon on when amdkfd and amd_iommu_v2 have been loaded, but
that would stall radeon.
Therefore, this patch moves the amdkfd part of GPU initialization to a
workqueue. When radeon calls amdkfd to perform GPU related initialization, it
will check if both amdkfd and amd_iommu_v2 have been loaded. If so, which is
the situation when the three drivers are compiled as modules, it will call the
relevant amdkfd function directly. If not, it will queue the initialization
work on the workqueue. The work function will schedule itself until both amdkfd
and amd_iommu_v2 have been loaded. Then, it will call the relevant amdkfd
function.
The workqueue is defined per kfd_dev structure (per GPU).
Signed-off-by: Oded Gabbay <[email protected]>
---
drivers/gpu/drm/amd/amdkfd/kfd_device.c | 72 +++++++++++++++++++++++++++++++--
drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 2 +
2 files changed, 70 insertions(+), 4 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_device.c b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
index 43884eb..cec5b4b 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device.c
@@ -24,6 +24,7 @@
#include <linux/bsearch.h>
#include <linux/pci.h>
#include <linux/slab.h>
+#include <linux/sched.h>
#include "kfd_priv.h"
#include "kfd_device_queue_manager.h"
@@ -40,6 +41,11 @@ struct kfd_deviceid {
const struct kfd_device_info *device_info;
};
+struct kfd_device_init_work {
+ struct work_struct kfd_work;
+ struct kfd_dev *dev;
+};
+
/* Please keep this sorted by increasing device id. */
static const struct kfd_deviceid supported_devices[] = {
{ 0x1304, &kaveri_device_info }, /* Kaveri */
@@ -99,6 +105,8 @@ struct kfd_dev *kgd2kfd_probe(struct kgd_dev *kgd, struct pci_dev *pdev)
kfd->pdev = pdev;
kfd->init_complete = false;
+ kfd->kfd_dev_wq = create_workqueue("kfd_dev_wq");
+
return kfd;
}
@@ -161,13 +169,10 @@ static void iommu_pasid_shutdown_callback(struct pci_dev *pdev, int pasid)
kfd_unbind_process_from_device(dev, pasid);
}
-bool kgd2kfd_device_init(struct kfd_dev *kfd,
- const struct kgd2kfd_shared_resources *gpu_resources)
+static bool kfd_kgd_device_init(struct kfd_dev *kfd)
{
unsigned int size;
- kfd->shared_resources = *gpu_resources;
-
/* calculate max size of mqds needed for queues */
size = max_num_of_processes *
max_num_of_queues_per_process *
@@ -249,6 +254,63 @@ out:
return kfd->init_complete;
}
+static void kfd_device_wq_init_device(struct work_struct *work)
+{
+ struct kfd_device_init_work *my_work;
+ struct kfd_dev *kfd;
+
+ my_work = (struct kfd_device_init_work *) work;
+
+ kfd = my_work->dev;
+
+ /*
+ * As long as amdkfd or amd_iommu_v2 are not initialized, we
+ * yield the processor
+ */
+ while ((!amdkfd_is_init_completed()) ||
+ (!amd_iommu_v2_is_init_completed()))
+ schedule();
+
+ kfd_kgd_device_init(kfd);
+}
+
+bool kgd2kfd_device_init(struct kfd_dev *kfd,
+ const struct kgd2kfd_shared_resources *gpu_resources)
+{
+ struct kfd_device_init_work *work;
+
+ kfd->shared_resources = *gpu_resources;
+
+ /*
+ * When amd_iommu_v2, amdkfd and radeon are compiled inside the kernel,
+ * there is no mechanism to enforce order of loading between the
+ * drivers. Therefore, we need to use an explicit form of
+ * synchronization to know when amdkfd and amd_iommu_v2 have finished
+ * there initialization routines
+ */
+ if ((!amdkfd_is_init_completed()) ||
+ (!amd_iommu_v2_is_init_completed())) {
+ BUG_ON(!kfd->kfd_dev_wq);
+
+ work = (struct kfd_device_init_work *)
+ kmalloc(sizeof(struct kfd_device_init_work),
+ GFP_ATOMIC);
+
+ if (!work) {
+ pr_err("kfd: no memory for device work queue\n");
+ return false;
+ }
+
+ INIT_WORK((struct work_struct *) work,
+ kfd_device_wq_init_device);
+ work->dev = kfd;
+ queue_work(kfd->kfd_dev_wq, (struct work_struct *) work);
+ return true;
+ }
+
+ return kfd_kgd_device_init(kfd);
+}
+
void kgd2kfd_device_exit(struct kfd_dev *kfd)
{
if (kfd->init_complete) {
@@ -258,6 +320,8 @@ void kgd2kfd_device_exit(struct kfd_dev *kfd)
kfd_topology_remove_device(kfd);
}
+ flush_workqueue(kfd->kfd_dev_wq);
+ destroy_workqueue(kfd->kfd_dev_wq);
kfree(kfd);
}
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
index 01df7e6..fc000a2 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h
@@ -151,6 +151,8 @@ struct kfd_dev {
* from the HW ring into a SW ring.
*/
bool interrupts_active;
+
+ struct workqueue_struct *kfd_dev_wq;
};
int amdkfd_is_init_completed(void);
--
2.1.0
When amdkfd and radeon are compiled inside the kernel image (not as modules),
radeon will load before amdkfd and will set *kfd2kgd to its interface
structure. Therefore, we must not set *kfd2kgd to NULL when amdkfd is loaded
because it will override radeon's initialization and cause kernel BUG.
Signed-off-by: Oded Gabbay <[email protected]>
---
drivers/gpu/drm/amd/amdkfd/kfd_module.c | 5 ++---
1 file changed, 2 insertions(+), 3 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_module.c b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
index 95d5af1..236562f 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_module.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
@@ -34,7 +34,7 @@
#define KFD_DRIVER_MINOR 7
#define KFD_DRIVER_PATCHLEVEL 0
-const struct kfd2kgd_calls *kfd2kgd;
+const struct kfd2kgd_calls *kfd2kgd = NULL;
static const struct kgd2kfd_calls kgd2kfd = {
.exit = kgd2kfd_exit,
.probe = kgd2kfd_probe,
@@ -84,14 +84,13 @@ EXPORT_SYMBOL(kgd2kfd_init);
void kgd2kfd_exit(void)
{
+ kfd2kgd = NULL;
}
static int __init kfd_module_init(void)
{
int err;
- kfd2kgd = NULL;
-
/* Verify module parameters */
if ((sched_policy < KFD_SCHED_POLICY_HWS) ||
(sched_policy > KFD_SCHED_POLICY_NO_HWS)) {
--
2.1.0
On Sat, Dec 20, 2014 at 10:46:12PM +0200, Oded Gabbay wrote:
> When amdkfd and radeon are compiled inside the kernel image (not as modules),
> radeon will load before amdkfd and will set *kfd2kgd to its interface
> structure. Therefore, we must not set *kfd2kgd to NULL when amdkfd is loaded
> because it will override radeon's initialization and cause kernel BUG.
>
> Signed-off-by: Oded Gabbay <[email protected]>
> ---
> drivers/gpu/drm/amd/amdkfd/kfd_module.c | 5 ++---
> 1 file changed, 2 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_module.c b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
> index 95d5af1..236562f 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_module.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
> @@ -34,7 +34,7 @@
> #define KFD_DRIVER_MINOR 7
> #define KFD_DRIVER_PATCHLEVEL 0
>
> -const struct kfd2kgd_calls *kfd2kgd;
> +const struct kfd2kgd_calls *kfd2kgd = NULL;
This change does nothing, that is what the original code did already.
thanks,
greg k-h
On 12/20/2014 11:25 PM, Greg KH wrote:
> On Sat, Dec 20, 2014 at 10:46:12PM +0200, Oded Gabbay wrote:
>> When amdkfd and radeon are compiled inside the kernel image (not as modules),
>> radeon will load before amdkfd and will set *kfd2kgd to its interface
>> structure. Therefore, we must not set *kfd2kgd to NULL when amdkfd is loaded
>> because it will override radeon's initialization and cause kernel BUG.
>>
>> Signed-off-by: Oded Gabbay <[email protected]>
>> ---
>> drivers/gpu/drm/amd/amdkfd/kfd_module.c | 5 ++---
>> 1 file changed, 2 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_module.c b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>> index 95d5af1..236562f 100644
>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>> @@ -34,7 +34,7 @@
>> #define KFD_DRIVER_MINOR 7
>> #define KFD_DRIVER_PATCHLEVEL 0
>>
>> -const struct kfd2kgd_calls *kfd2kgd;
>> +const struct kfd2kgd_calls *kfd2kgd = NULL;
>
> This change does nothing, that is what the original code did already.
>
> thanks,
>
> greg k-h
Yeah. Sorry, I will remove it.
Oded
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.tux.org/lkml/
>
On 12/21/2014 01:27 PM, Christian König wrote:
> Am 20.12.2014 um 21:46 schrieb Oded Gabbay:
>> When amdkfd and radeon are compiled inside the kernel image (not as modules),
>> radeon will load before amdkfd and will set *kfd2kgd to its interface
>> structure. Therefore, we must not set *kfd2kgd to NULL when amdkfd is loaded
>> because it will override radeon's initialization and cause kernel BUG.
>>
>> Signed-off-by: Oded Gabbay <[email protected]>
>
> You should probably rather fix the dependency between the two modules to get an
> determined load order instead of doing such nasty workarounds.
>
> Christian.
The problem is that when modules are compiled inside the kernel, there is NO
determined load order and there is no mechanism to enforce that. If there is/was
such a mechanism, I would of course prefer to use it.
Actually, I don't understand why the kernel doesn't enforce the order according
to the use of exported symbols, like it does with modules.
There will always be dependencies between kgd (radeon) and amdkfd and between
amdkfd and amd_iommu_v2. I don't think I can eliminate those dependencies, not
without a very complex solution. And the fact that this complex solution occurs
only in a very specific use case (all modules compiled in), makes me less
inclined to do that.
So I don't see it as a "nasty workaround". I would call it just a "workaround"
for a specific use case, which should be solved by a generic solution to the
kernel enforcing load orders.
Oded
>
>> ---
>> drivers/gpu/drm/amd/amdkfd/kfd_module.c | 5 ++---
>> 1 file changed, 2 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>> b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>> index 95d5af1..236562f 100644
>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>> @@ -34,7 +34,7 @@
>> #define KFD_DRIVER_MINOR 7
>> #define KFD_DRIVER_PATCHLEVEL 0
>> -const struct kfd2kgd_calls *kfd2kgd;
>> +const struct kfd2kgd_calls *kfd2kgd = NULL;
>> static const struct kgd2kfd_calls kgd2kfd = {
>> .exit = kgd2kfd_exit,
>> .probe = kgd2kfd_probe,
>> @@ -84,14 +84,13 @@ EXPORT_SYMBOL(kgd2kfd_init);
>> void kgd2kfd_exit(void)
>> {
>> + kfd2kgd = NULL;
>> }
>> static int __init kfd_module_init(void)
>> {
>> int err;
>> - kfd2kgd = NULL;
>> -
>> /* Verify module parameters */
>> if ((sched_policy < KFD_SCHED_POLICY_HWS) ||
>> (sched_policy > KFD_SCHED_POLICY_NO_HWS)) {
>
Am 20.12.2014 um 21:46 schrieb Oded Gabbay:
> When amdkfd and radeon are compiled inside the kernel image (not as modules),
> radeon will load before amdkfd and will set *kfd2kgd to its interface
> structure. Therefore, we must not set *kfd2kgd to NULL when amdkfd is loaded
> because it will override radeon's initialization and cause kernel BUG.
>
> Signed-off-by: Oded Gabbay <[email protected]>
You should probably rather fix the dependency between the two modules to
get an determined load order instead of doing such nasty workarounds.
Christian.
> ---
> drivers/gpu/drm/amd/amdkfd/kfd_module.c | 5 ++---
> 1 file changed, 2 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_module.c b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
> index 95d5af1..236562f 100644
> --- a/drivers/gpu/drm/amd/amdkfd/kfd_module.c
> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
> @@ -34,7 +34,7 @@
> #define KFD_DRIVER_MINOR 7
> #define KFD_DRIVER_PATCHLEVEL 0
>
> -const struct kfd2kgd_calls *kfd2kgd;
> +const struct kfd2kgd_calls *kfd2kgd = NULL;
> static const struct kgd2kfd_calls kgd2kfd = {
> .exit = kgd2kfd_exit,
> .probe = kgd2kfd_probe,
> @@ -84,14 +84,13 @@ EXPORT_SYMBOL(kgd2kfd_init);
>
> void kgd2kfd_exit(void)
> {
> + kfd2kgd = NULL;
> }
>
> static int __init kfd_module_init(void)
> {
> int err;
>
> - kfd2kgd = NULL;
> -
> /* Verify module parameters */
> if ((sched_policy < KFD_SCHED_POLICY_HWS) ||
> (sched_policy > KFD_SCHED_POLICY_NO_HWS)) {
Am 21.12.2014 um 12:34 schrieb Oded Gabbay:
>
>
> On 12/21/2014 01:27 PM, Christian König wrote:
>> Am 20.12.2014 um 21:46 schrieb Oded Gabbay:
>>> When amdkfd and radeon are compiled inside the kernel image (not as
>>> modules),
>>> radeon will load before amdkfd and will set *kfd2kgd to its interface
>>> structure. Therefore, we must not set *kfd2kgd to NULL when amdkfd
>>> is loaded
>>> because it will override radeon's initialization and cause kernel BUG.
>>>
>>> Signed-off-by: Oded Gabbay <[email protected]>
>>
>> You should probably rather fix the dependency between the two modules
>> to get an
>> determined load order instead of doing such nasty workarounds.
>>
>> Christian.
>
> The problem is that when modules are compiled inside the kernel, there
> is NO determined load order and there is no mechanism to enforce that.
> If there is/was such a mechanism, I would of course prefer to use it.
There should be an determined order based on the symbol use, otherwise
initializing most of the kernel modules wouldn't work as expected. For
example radeon depends on the drm module must be loaded before radeon is
loaded.
>
> Actually, I don't understand why the kernel doesn't enforce the order
> according to the use of exported symbols, like it does with modules.
Yeah, that's indeed rather strange. There must be something in the
amdkfd code which broke that somehow.
As far as I understand you the desired init order is radeon and
amd_iommu_v2 first and then amdkfd, right? So what happens when you boot
with radeon, amd_iommu_v2 and amdkfd blacklisted for automatically load
and only load amdkfd manually?
> There will always be dependencies between kgd (radeon) and amdkfd and
> between amdkfd and amd_iommu_v2. I don't think I can eliminate those
> dependencies, not without a very complex solution. And the fact that
> this complex solution occurs only in a very specific use case (all
> modules compiled in), makes me less inclined to do that.
>
> So I don't see it as a "nasty workaround". I would call it just a
> "workaround" for a specific use case, which should be solved by a
> generic solution to the kernel enforcing load orders.
The normal kernel module handling already should provide the correct
init order, so I would still call this a rather nasty workaround because
we couldn't find the underlying problem.
Christian.
>
> Oded
>>
>>> ---
>>> drivers/gpu/drm/amd/amdkfd/kfd_module.c | 5 ++---
>>> 1 file changed, 2 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>> b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>> index 95d5af1..236562f 100644
>>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>> @@ -34,7 +34,7 @@
>>> #define KFD_DRIVER_MINOR 7
>>> #define KFD_DRIVER_PATCHLEVEL 0
>>> -const struct kfd2kgd_calls *kfd2kgd;
>>> +const struct kfd2kgd_calls *kfd2kgd = NULL;
>>> static const struct kgd2kfd_calls kgd2kfd = {
>>> .exit = kgd2kfd_exit,
>>> .probe = kgd2kfd_probe,
>>> @@ -84,14 +84,13 @@ EXPORT_SYMBOL(kgd2kfd_init);
>>> void kgd2kfd_exit(void)
>>> {
>>> + kfd2kgd = NULL;
>>> }
>>> static int __init kfd_module_init(void)
>>> {
>>> int err;
>>> - kfd2kgd = NULL;
>>> -
>>> /* Verify module parameters */
>>> if ((sched_policy < KFD_SCHED_POLICY_HWS) ||
>>> (sched_policy > KFD_SCHED_POLICY_NO_HWS)) {
>>
On 12/21/2014 02:19 PM, Christian König wrote:
> Am 21.12.2014 um 12:34 schrieb Oded Gabbay:
>>
>>
>> On 12/21/2014 01:27 PM, Christian König wrote:
>>> Am 20.12.2014 um 21:46 schrieb Oded Gabbay:
>>>> When amdkfd and radeon are compiled inside the kernel image (not as modules),
>>>> radeon will load before amdkfd and will set *kfd2kgd to its interface
>>>> structure. Therefore, we must not set *kfd2kgd to NULL when amdkfd is loaded
>>>> because it will override radeon's initialization and cause kernel BUG.
>>>>
>>>> Signed-off-by: Oded Gabbay <[email protected]>
>>>
>>> You should probably rather fix the dependency between the two modules to get an
>>> determined load order instead of doing such nasty workarounds.
>>>
>>> Christian.
>>
>> The problem is that when modules are compiled inside the kernel, there is NO
>> determined load order and there is no mechanism to enforce that. If there
>> is/was such a mechanism, I would of course prefer to use it.
>
> There should be an determined order based on the symbol use, otherwise
> initializing most of the kernel modules wouldn't work as expected. For example
> radeon depends on the drm module must be loaded before radeon is loaded.
There should be, but when the modules are compiled in, they are loaded based on
link order only, if they are in the same group, and the groups are loaded by a
pre-defined order.
The groups are: pure, core, postcore, arch, subsys, fs, device (which represents
all the modules) and late. See init.h
So radeon, amdkfd and amd_iommu_v2 are all in device group, and in the group
they are ordered by their link order.
Yes, radeon loads after drm, because drm*.o are before radeon*.o in the
Makefile. See
http://stackoverflow.com/questions/5669647/linux-order-of-statically-linked-module-loading
>
>>
>> Actually, I don't understand why the kernel doesn't enforce the order
>> according to the use of exported symbols, like it does with modules.
>
> Yeah, that's indeed rather strange. There must be something in the amdkfd code
> which broke that somehow.
IMO, that's a far-fetched guess. Could you point to something more specific ?
>
> As far as I understand you the desired init order is radeon and amd_iommu_v2
> first and then amdkfd, right?
Actually no. The preferred order is amd_iommu_v2, amdkfd and radeon last. This
is the order that happens when all three are built as modules. More accurately,
radeon inits, but its init triggers amdkfd init, which triggers amd_iommu_v2
init. So before radeon reaches its probe stage, all the modules were initialized.
So what happens when you boot with radeon,
> amd_iommu_v2 and amdkfd blacklisted for automatically load and only load amdkfd
> manually?
As said above, that's ok.
>
>> There will always be dependencies between kgd (radeon) and amdkfd and between
>> amdkfd and amd_iommu_v2. I don't think I can eliminate those dependencies, not
>> without a very complex solution. And the fact that this complex solution
>> occurs only in a very specific use case (all modules compiled in), makes me
>> less inclined to do that.
>>
>> So I don't see it as a "nasty workaround". I would call it just a "workaround"
>> for a specific use case, which should be solved by a generic solution to the
>> kernel enforcing load orders.
>
> The normal kernel module handling already should provide the correct init order,
> so I would still call this a rather nasty workaround because we couldn't find
> the underlying problem.
Well, the normal kernel module handling doesn't work when all modules are
compiled in. I'm not a huge expert on this issue so I had Joerg Roedel help me
analyze this (thanks Joerg) and he agreed that there is no enforcement of order
in this case.
>
> Christian.
>
>>
>> Oded
>>>
>>>> ---
>>>> drivers/gpu/drm/amd/amdkfd/kfd_module.c | 5 ++---
>>>> 1 file changed, 2 insertions(+), 3 deletions(-)
>>>>
>>>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>>> b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>>> index 95d5af1..236562f 100644
>>>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>>> @@ -34,7 +34,7 @@
>>>> #define KFD_DRIVER_MINOR 7
>>>> #define KFD_DRIVER_PATCHLEVEL 0
>>>> -const struct kfd2kgd_calls *kfd2kgd;
>>>> +const struct kfd2kgd_calls *kfd2kgd = NULL;
>>>> static const struct kgd2kfd_calls kgd2kfd = {
>>>> .exit = kgd2kfd_exit,
>>>> .probe = kgd2kfd_probe,
>>>> @@ -84,14 +84,13 @@ EXPORT_SYMBOL(kgd2kfd_init);
>>>> void kgd2kfd_exit(void)
>>>> {
>>>> + kfd2kgd = NULL;
>>>> }
>>>> static int __init kfd_module_init(void)
>>>> {
>>>> int err;
>>>> - kfd2kgd = NULL;
>>>> -
>>>> /* Verify module parameters */
>>>> if ((sched_policy < KFD_SCHED_POLICY_HWS) ||
>>>> (sched_policy > KFD_SCHED_POLICY_NO_HWS)) {
>>>
>
On 12/21/2014 03:06 PM, Oded Gabbay wrote:
>
>
> On 12/21/2014 02:19 PM, Christian König wrote:
>> Am 21.12.2014 um 12:34 schrieb Oded Gabbay:
>>>
>>>
>>> On 12/21/2014 01:27 PM, Christian König wrote:
>>>> Am 20.12.2014 um 21:46 schrieb Oded Gabbay:
>>>>> When amdkfd and radeon are compiled inside the kernel image (not as modules),
>>>>> radeon will load before amdkfd and will set *kfd2kgd to its interface
>>>>> structure. Therefore, we must not set *kfd2kgd to NULL when amdkfd is loaded
>>>>> because it will override radeon's initialization and cause kernel BUG.
>>>>>
>>>>> Signed-off-by: Oded Gabbay <[email protected]>
>>>>
>>>> You should probably rather fix the dependency between the two modules to get an
>>>> determined load order instead of doing such nasty workarounds.
>>>>
>>>> Christian.
>>>
>>> The problem is that when modules are compiled inside the kernel, there is NO
>>> determined load order and there is no mechanism to enforce that. If there
>>> is/was such a mechanism, I would of course prefer to use it.
>>
>> There should be an determined order based on the symbol use, otherwise
>> initializing most of the kernel modules wouldn't work as expected. For example
>> radeon depends on the drm module must be loaded before radeon is loaded.
> There should be, but when the modules are compiled in, they are loaded based on
> link order only, if they are in the same group, and the groups are loaded by a
> pre-defined order.
> The groups are: pure, core, postcore, arch, subsys, fs, device (which represents
> all the modules) and late. See init.h
>
> So radeon, amdkfd and amd_iommu_v2 are all in device group, and in the group
> they are ordered by their link order.
>
> Yes, radeon loads after drm, because drm*.o are before radeon*.o in the
> Makefile. See
> http://stackoverflow.com/questions/5669647/linux-order-of-statically-linked-module-loading
>
So I tried moving amdkfd inside the Makefile before radeon, and that made amdkfd
load before radeon did. This solves my first problem - order between amdkfd and
radeon. However, amd_iommu_v2 doesn't belong to the drm Makefile, and I don't
want to move iommu before gpu, so I don't have a solution for the order between
amdkfd and amd_iommu_v2.
Oded
>
>
>>
>>>
>>> Actually, I don't understand why the kernel doesn't enforce the order
>>> according to the use of exported symbols, like it does with modules.
>>
>> Yeah, that's indeed rather strange. There must be something in the amdkfd code
>> which broke that somehow.
> IMO, that's a far-fetched guess. Could you point to something more specific ?
>
>>
>> As far as I understand you the desired init order is radeon and amd_iommu_v2
>> first and then amdkfd, right?
> Actually no. The preferred order is amd_iommu_v2, amdkfd and radeon last. This
> is the order that happens when all three are built as modules. More accurately,
> radeon inits, but its init triggers amdkfd init, which triggers amd_iommu_v2
> init. So before radeon reaches its probe stage, all the modules were initialized.
>
> So what happens when you boot with radeon,
>> amd_iommu_v2 and amdkfd blacklisted for automatically load and only load amdkfd
>> manually?
> As said above, that's ok.
>>
>>> There will always be dependencies between kgd (radeon) and amdkfd and between
>>> amdkfd and amd_iommu_v2. I don't think I can eliminate those dependencies, not
>>> without a very complex solution. And the fact that this complex solution
>>> occurs only in a very specific use case (all modules compiled in), makes me
>>> less inclined to do that.
>>>
>>> So I don't see it as a "nasty workaround". I would call it just a "workaround"
>>> for a specific use case, which should be solved by a generic solution to the
>>> kernel enforcing load orders.
>>
>> The normal kernel module handling already should provide the correct init order,
>> so I would still call this a rather nasty workaround because we couldn't find
>> the underlying problem.
> Well, the normal kernel module handling doesn't work when all modules are
> compiled in. I'm not a huge expert on this issue so I had Joerg Roedel help me
> analyze this (thanks Joerg) and he agreed that there is no enforcement of order
> in this case.
>
>>
>> Christian.
>>
>>>
>>> Oded
>>>>
>>>>> ---
>>>>> drivers/gpu/drm/amd/amdkfd/kfd_module.c | 5 ++---
>>>>> 1 file changed, 2 insertions(+), 3 deletions(-)
>>>>>
>>>>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>>>> b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>>>> index 95d5af1..236562f 100644
>>>>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>>>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>>>> @@ -34,7 +34,7 @@
>>>>> #define KFD_DRIVER_MINOR 7
>>>>> #define KFD_DRIVER_PATCHLEVEL 0
>>>>> -const struct kfd2kgd_calls *kfd2kgd;
>>>>> +const struct kfd2kgd_calls *kfd2kgd = NULL;
>>>>> static const struct kgd2kfd_calls kgd2kfd = {
>>>>> .exit = kgd2kfd_exit,
>>>>> .probe = kgd2kfd_probe,
>>>>> @@ -84,14 +84,13 @@ EXPORT_SYMBOL(kgd2kfd_init);
>>>>> void kgd2kfd_exit(void)
>>>>> {
>>>>> + kfd2kgd = NULL;
>>>>> }
>>>>> static int __init kfd_module_init(void)
>>>>> {
>>>>> int err;
>>>>> - kfd2kgd = NULL;
>>>>> -
>>>>> /* Verify module parameters */
>>>>> if ((sched_policy < KFD_SCHED_POLICY_HWS) ||
>>>>> (sched_policy > KFD_SCHED_POLICY_NO_HWS)) {
>>>>
>>
> _______________________________________________
> dri-devel mailing list
> [email protected]
> http://lists.freedesktop.org/mailman/listinfo/dri-devel
> There should be, but when the modules are compiled in, they are loaded
> based on
> link order only, if they are in the same group, and the groups are
> loaded by a
> pre-defined order.
Is that really still up to date? I've seen effort to change that
something like 10+ years ago when Rusty reworked the module system. And
it is comming up on the lists again from time to time.
> I don't want to move iommu before gpu, so I don't have a solution for
> the order between amdkfd and amd_iommu_v2.
Why not? That's still better than creating a kernel workqueue,
scheduling one work item on it, rescheduling the task until everything
is completed and you can continue.
Christian.
Am 21.12.2014 um 14:24 schrieb Oded Gabbay:
>
>
> On 12/21/2014 03:06 PM, Oded Gabbay wrote:
>>
>>
>> On 12/21/2014 02:19 PM, Christian König wrote:
>>> Am 21.12.2014 um 12:34 schrieb Oded Gabbay:
>>>>
>>>>
>>>> On 12/21/2014 01:27 PM, Christian König wrote:
>>>>> Am 20.12.2014 um 21:46 schrieb Oded Gabbay:
>>>>>> When amdkfd and radeon are compiled inside the kernel image (not
>>>>>> as modules),
>>>>>> radeon will load before amdkfd and will set *kfd2kgd to its
>>>>>> interface
>>>>>> structure. Therefore, we must not set *kfd2kgd to NULL when
>>>>>> amdkfd is loaded
>>>>>> because it will override radeon's initialization and cause kernel
>>>>>> BUG.
>>>>>>
>>>>>> Signed-off-by: Oded Gabbay <[email protected]>
>>>>>
>>>>> You should probably rather fix the dependency between the two
>>>>> modules to get an
>>>>> determined load order instead of doing such nasty workarounds.
>>>>>
>>>>> Christian.
>>>>
>>>> The problem is that when modules are compiled inside the kernel,
>>>> there is NO
>>>> determined load order and there is no mechanism to enforce that. If
>>>> there
>>>> is/was such a mechanism, I would of course prefer to use it.
>>>
>>> There should be an determined order based on the symbol use, otherwise
>>> initializing most of the kernel modules wouldn't work as expected.
>>> For example
>>> radeon depends on the drm module must be loaded before radeon is
>>> loaded.
>> There should be, but when the modules are compiled in, they are
>> loaded based on
>> link order only, if they are in the same group, and the groups are
>> loaded by a
>> pre-defined order.
>> The groups are: pure, core, postcore, arch, subsys, fs, device (which
>> represents
>> all the modules) and late. See init.h
>>
>> So radeon, amdkfd and amd_iommu_v2 are all in device group, and in
>> the group
>> they are ordered by their link order.
>>
>> Yes, radeon loads after drm, because drm*.o are before radeon*.o in the
>> Makefile. See
>> http://stackoverflow.com/questions/5669647/linux-order-of-statically-linked-module-loading
>>
>>
>
> So I tried moving amdkfd inside the Makefile before radeon, and that
> made amdkfd load before radeon did. This solves my first problem -
> order between amdkfd and radeon. However, amd_iommu_v2 doesn't belong
> to the drm Makefile, and I don't want to move iommu before gpu, so I
> don't have a solution for the order between amdkfd and amd_iommu_v2.
>
> Oded
>>
>>
>>>
>>>>
>>>> Actually, I don't understand why the kernel doesn't enforce the order
>>>> according to the use of exported symbols, like it does with modules.
>>>
>>> Yeah, that's indeed rather strange. There must be something in the
>>> amdkfd code
>>> which broke that somehow.
>> IMO, that's a far-fetched guess. Could you point to something more
>> specific ?
>>
>>>
>>> As far as I understand you the desired init order is radeon and
>>> amd_iommu_v2
>>> first and then amdkfd, right?
>> Actually no. The preferred order is amd_iommu_v2, amdkfd and radeon
>> last. This
>> is the order that happens when all three are built as modules. More
>> accurately,
>> radeon inits, but its init triggers amdkfd init, which triggers
>> amd_iommu_v2
>> init. So before radeon reaches its probe stage, all the modules were
>> initialized.
>>
>> So what happens when you boot with radeon,
>>> amd_iommu_v2 and amdkfd blacklisted for automatically load and only
>>> load amdkfd
>>> manually?
>> As said above, that's ok.
>>>
>>>> There will always be dependencies between kgd (radeon) and amdkfd
>>>> and between
>>>> amdkfd and amd_iommu_v2. I don't think I can eliminate those
>>>> dependencies, not
>>>> without a very complex solution. And the fact that this complex
>>>> solution
>>>> occurs only in a very specific use case (all modules compiled in),
>>>> makes me
>>>> less inclined to do that.
>>>>
>>>> So I don't see it as a "nasty workaround". I would call it just a
>>>> "workaround"
>>>> for a specific use case, which should be solved by a generic
>>>> solution to the
>>>> kernel enforcing load orders.
>>>
>>> The normal kernel module handling already should provide the correct
>>> init order,
>>> so I would still call this a rather nasty workaround because we
>>> couldn't find
>>> the underlying problem.
>> Well, the normal kernel module handling doesn't work when all modules
>> are
>> compiled in. I'm not a huge expert on this issue so I had Joerg
>> Roedel help me
>> analyze this (thanks Joerg) and he agreed that there is no
>> enforcement of order
>> in this case.
>>
>>>
>>> Christian.
>>>
>>>>
>>>> Oded
>>>>>
>>>>>> ---
>>>>>> drivers/gpu/drm/amd/amdkfd/kfd_module.c | 5 ++---
>>>>>> 1 file changed, 2 insertions(+), 3 deletions(-)
>>>>>>
>>>>>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>>>>> b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>>>>> index 95d5af1..236562f 100644
>>>>>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>>>>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>>>>> @@ -34,7 +34,7 @@
>>>>>> #define KFD_DRIVER_MINOR 7
>>>>>> #define KFD_DRIVER_PATCHLEVEL 0
>>>>>> -const struct kfd2kgd_calls *kfd2kgd;
>>>>>> +const struct kfd2kgd_calls *kfd2kgd = NULL;
>>>>>> static const struct kgd2kfd_calls kgd2kfd = {
>>>>>> .exit = kgd2kfd_exit,
>>>>>> .probe = kgd2kfd_probe,
>>>>>> @@ -84,14 +84,13 @@ EXPORT_SYMBOL(kgd2kfd_init);
>>>>>> void kgd2kfd_exit(void)
>>>>>> {
>>>>>> + kfd2kgd = NULL;
>>>>>> }
>>>>>> static int __init kfd_module_init(void)
>>>>>> {
>>>>>> int err;
>>>>>> - kfd2kgd = NULL;
>>>>>> -
>>>>>> /* Verify module parameters */
>>>>>> if ((sched_policy < KFD_SCHED_POLICY_HWS) ||
>>>>>> (sched_policy > KFD_SCHED_POLICY_NO_HWS)) {
>>>>>
>>>
>> _______________________________________________
>> dri-devel mailing list
>> [email protected]
>> http://lists.freedesktop.org/mailman/listinfo/dri-devel
On 12/21/2014 05:57 PM, Christian König wrote:
>> There should be, but when the modules are compiled in, they are loaded based on
>> link order only, if they are in the same group, and the groups are loaded by a
>> pre-defined order.
> Is that really still up to date? I've seen effort to change that something like
> 10+ years ago when Rusty reworked the module system. And it is comming up on the
> lists again from time to time.
From what I can see in the Makefile rules, code and google, yes, that's still
the situation. If someone will prove me wrong I will be more than happy to
correct my code.
>
>> I don't want to move iommu before gpu, so I don't have a solution for the
>> order between amdkfd and amd_iommu_v2.
> Why not? That's still better than creating a kernel workqueue, scheduling one
> work item on it, rescheduling the task until everything is completed and you can
> continue.
Because I don't know the consequences of moving an entire subsystem in front of
another one. In addition, even if everyone agrees, I'm pretty sure that Linus
won't be happy to do that in -rc stages. So maybe this is something to consider
for 3.20 merge window, but I would still like to provide a solution for 3.19.
Oded
>
> Christian.
>
> Am 21.12.2014 um 14:24 schrieb Oded Gabbay:
>>
>>
>> On 12/21/2014 03:06 PM, Oded Gabbay wrote:
>>>
>>>
>>> On 12/21/2014 02:19 PM, Christian König wrote:
>>>> Am 21.12.2014 um 12:34 schrieb Oded Gabbay:
>>>>>
>>>>>
>>>>> On 12/21/2014 01:27 PM, Christian König wrote:
>>>>>> Am 20.12.2014 um 21:46 schrieb Oded Gabbay:
>>>>>>> When amdkfd and radeon are compiled inside the kernel image (not as
>>>>>>> modules),
>>>>>>> radeon will load before amdkfd and will set *kfd2kgd to its interface
>>>>>>> structure. Therefore, we must not set *kfd2kgd to NULL when amdkfd is loaded
>>>>>>> because it will override radeon's initialization and cause kernel BUG.
>>>>>>>
>>>>>>> Signed-off-by: Oded Gabbay <[email protected]>
>>>>>>
>>>>>> You should probably rather fix the dependency between the two modules to
>>>>>> get an
>>>>>> determined load order instead of doing such nasty workarounds.
>>>>>>
>>>>>> Christian.
>>>>>
>>>>> The problem is that when modules are compiled inside the kernel, there is NO
>>>>> determined load order and there is no mechanism to enforce that. If there
>>>>> is/was such a mechanism, I would of course prefer to use it.
>>>>
>>>> There should be an determined order based on the symbol use, otherwise
>>>> initializing most of the kernel modules wouldn't work as expected. For example
>>>> radeon depends on the drm module must be loaded before radeon is loaded.
>>> There should be, but when the modules are compiled in, they are loaded based on
>>> link order only, if they are in the same group, and the groups are loaded by a
>>> pre-defined order.
>>> The groups are: pure, core, postcore, arch, subsys, fs, device (which represents
>>> all the modules) and late. See init.h
>>>
>>> So radeon, amdkfd and amd_iommu_v2 are all in device group, and in the group
>>> they are ordered by their link order.
>>>
>>> Yes, radeon loads after drm, because drm*.o are before radeon*.o in the
>>> Makefile. See
>>> http://stackoverflow.com/questions/5669647/linux-order-of-statically-linked-module-loading
>>>
>>>
>>
>> So I tried moving amdkfd inside the Makefile before radeon, and that made
>> amdkfd load before radeon did. This solves my first problem - order between
>> amdkfd and radeon. However, amd_iommu_v2 doesn't belong to the drm Makefile,
>> and I don't want to move iommu before gpu, so I don't have a solution for the
>> order between amdkfd and amd_iommu_v2.
>>
>> Oded
>>>
>>>
>>>>
>>>>>
>>>>> Actually, I don't understand why the kernel doesn't enforce the order
>>>>> according to the use of exported symbols, like it does with modules.
>>>>
>>>> Yeah, that's indeed rather strange. There must be something in the amdkfd code
>>>> which broke that somehow.
>>> IMO, that's a far-fetched guess. Could you point to something more specific ?
>>>
>>>>
>>>> As far as I understand you the desired init order is radeon and amd_iommu_v2
>>>> first and then amdkfd, right?
>>> Actually no. The preferred order is amd_iommu_v2, amdkfd and radeon last. This
>>> is the order that happens when all three are built as modules. More accurately,
>>> radeon inits, but its init triggers amdkfd init, which triggers amd_iommu_v2
>>> init. So before radeon reaches its probe stage, all the modules were
>>> initialized.
>>>
>>> So what happens when you boot with radeon,
>>>> amd_iommu_v2 and amdkfd blacklisted for automatically load and only load amdkfd
>>>> manually?
>>> As said above, that's ok.
>>>>
>>>>> There will always be dependencies between kgd (radeon) and amdkfd and between
>>>>> amdkfd and amd_iommu_v2. I don't think I can eliminate those dependencies, not
>>>>> without a very complex solution. And the fact that this complex solution
>>>>> occurs only in a very specific use case (all modules compiled in), makes me
>>>>> less inclined to do that.
>>>>>
>>>>> So I don't see it as a "nasty workaround". I would call it just a "workaround"
>>>>> for a specific use case, which should be solved by a generic solution to the
>>>>> kernel enforcing load orders.
>>>>
>>>> The normal kernel module handling already should provide the correct init
>>>> order,
>>>> so I would still call this a rather nasty workaround because we couldn't find
>>>> the underlying problem.
>>> Well, the normal kernel module handling doesn't work when all modules are
>>> compiled in. I'm not a huge expert on this issue so I had Joerg Roedel help me
>>> analyze this (thanks Joerg) and he agreed that there is no enforcement of order
>>> in this case.
>>>
>>>>
>>>> Christian.
>>>>
>>>>>
>>>>> Oded
>>>>>>
>>>>>>> ---
>>>>>>> drivers/gpu/drm/amd/amdkfd/kfd_module.c | 5 ++---
>>>>>>> 1 file changed, 2 insertions(+), 3 deletions(-)
>>>>>>>
>>>>>>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>>>>>> b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>>>>>> index 95d5af1..236562f 100644
>>>>>>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>>>>>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>>>>>> @@ -34,7 +34,7 @@
>>>>>>> #define KFD_DRIVER_MINOR 7
>>>>>>> #define KFD_DRIVER_PATCHLEVEL 0
>>>>>>> -const struct kfd2kgd_calls *kfd2kgd;
>>>>>>> +const struct kfd2kgd_calls *kfd2kgd = NULL;
>>>>>>> static const struct kgd2kfd_calls kgd2kfd = {
>>>>>>> .exit = kgd2kfd_exit,
>>>>>>> .probe = kgd2kfd_probe,
>>>>>>> @@ -84,14 +84,13 @@ EXPORT_SYMBOL(kgd2kfd_init);
>>>>>>> void kgd2kfd_exit(void)
>>>>>>> {
>>>>>>> + kfd2kgd = NULL;
>>>>>>> }
>>>>>>> static int __init kfd_module_init(void)
>>>>>>> {
>>>>>>> int err;
>>>>>>> - kfd2kgd = NULL;
>>>>>>> -
>>>>>>> /* Verify module parameters */
>>>>>>> if ((sched_policy < KFD_SCHED_POLICY_HWS) ||
>>>>>>> (sched_policy > KFD_SCHED_POLICY_NO_HWS)) {
>>>>>>
>>>>
>>> _______________________________________________
>>> dri-devel mailing list
>>> [email protected]
>>> http://lists.freedesktop.org/mailman/listinfo/dri-devel
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.tux.org/lkml/
Am 21.12.2014 um 17:03 schrieb Oded Gabbay:
>
>
> On 12/21/2014 05:57 PM, Christian König wrote:
>>> There should be, but when the modules are compiled in, they are
>>> loaded based on
>>> link order only, if they are in the same group, and the groups are
>>> loaded by a
>>> pre-defined order.
>> Is that really still up to date? I've seen effort to change that
>> something like
>> 10+ years ago when Rusty reworked the module system. And it is
>> comming up on the
>> lists again from time to time.
> From what I can see in the Makefile rules, code and google, yes,
> that's still the situation. If someone will prove me wrong I will be
> more than happy to correct my code.
>>
>>> I don't want to move iommu before gpu, so I don't have a solution
>>> for the
>>> order between amdkfd and amd_iommu_v2.
>> Why not? That's still better than creating a kernel workqueue,
>> scheduling one
>> work item on it, rescheduling the task until everything is completed
>> and you can
>> continue.
> Because I don't know the consequences of moving an entire subsystem in
> front of another one. In addition, even if everyone agrees, I'm pretty
> sure that Linus won't be happy to do that in -rc stages. So maybe this
> is something to consider for 3.20 merge window, but I would still like
> to provide a solution for 3.19.
Yeah, true indeed. How about depending on everything being compiled as
module for 3.19 then? Still better than having such a hack in the driver
for as a temporary workaround for one release.
Christian.
>
> Oded
>>
>> Christian.
>>
>> Am 21.12.2014 um 14:24 schrieb Oded Gabbay:
>>>
>>>
>>> On 12/21/2014 03:06 PM, Oded Gabbay wrote:
>>>>
>>>>
>>>> On 12/21/2014 02:19 PM, Christian König wrote:
>>>>> Am 21.12.2014 um 12:34 schrieb Oded Gabbay:
>>>>>>
>>>>>>
>>>>>> On 12/21/2014 01:27 PM, Christian König wrote:
>>>>>>> Am 20.12.2014 um 21:46 schrieb Oded Gabbay:
>>>>>>>> When amdkfd and radeon are compiled inside the kernel image
>>>>>>>> (not as
>>>>>>>> modules),
>>>>>>>> radeon will load before amdkfd and will set *kfd2kgd to its
>>>>>>>> interface
>>>>>>>> structure. Therefore, we must not set *kfd2kgd to NULL when
>>>>>>>> amdkfd is loaded
>>>>>>>> because it will override radeon's initialization and cause
>>>>>>>> kernel BUG.
>>>>>>>>
>>>>>>>> Signed-off-by: Oded Gabbay <[email protected]>
>>>>>>>
>>>>>>> You should probably rather fix the dependency between the two
>>>>>>> modules to
>>>>>>> get an
>>>>>>> determined load order instead of doing such nasty workarounds.
>>>>>>>
>>>>>>> Christian.
>>>>>>
>>>>>> The problem is that when modules are compiled inside the kernel,
>>>>>> there is NO
>>>>>> determined load order and there is no mechanism to enforce that.
>>>>>> If there
>>>>>> is/was such a mechanism, I would of course prefer to use it.
>>>>>
>>>>> There should be an determined order based on the symbol use,
>>>>> otherwise
>>>>> initializing most of the kernel modules wouldn't work as expected.
>>>>> For example
>>>>> radeon depends on the drm module must be loaded before radeon is
>>>>> loaded.
>>>> There should be, but when the modules are compiled in, they are
>>>> loaded based on
>>>> link order only, if they are in the same group, and the groups are
>>>> loaded by a
>>>> pre-defined order.
>>>> The groups are: pure, core, postcore, arch, subsys, fs, device
>>>> (which represents
>>>> all the modules) and late. See init.h
>>>>
>>>> So radeon, amdkfd and amd_iommu_v2 are all in device group, and in
>>>> the group
>>>> they are ordered by their link order.
>>>>
>>>> Yes, radeon loads after drm, because drm*.o are before radeon*.o in
>>>> the
>>>> Makefile. See
>>>> http://stackoverflow.com/questions/5669647/linux-order-of-statically-linked-module-loading
>>>>
>>>>
>>>>
>>>
>>> So I tried moving amdkfd inside the Makefile before radeon, and that
>>> made
>>> amdkfd load before radeon did. This solves my first problem - order
>>> between
>>> amdkfd and radeon. However, amd_iommu_v2 doesn't belong to the drm
>>> Makefile,
>>> and I don't want to move iommu before gpu, so I don't have a
>>> solution for the
>>> order between amdkfd and amd_iommu_v2.
>>>
>>> Oded
>>>>
>>>>
>>>>>
>>>>>>
>>>>>> Actually, I don't understand why the kernel doesn't enforce the
>>>>>> order
>>>>>> according to the use of exported symbols, like it does with modules.
>>>>>
>>>>> Yeah, that's indeed rather strange. There must be something in the
>>>>> amdkfd code
>>>>> which broke that somehow.
>>>> IMO, that's a far-fetched guess. Could you point to something more
>>>> specific ?
>>>>
>>>>>
>>>>> As far as I understand you the desired init order is radeon and
>>>>> amd_iommu_v2
>>>>> first and then amdkfd, right?
>>>> Actually no. The preferred order is amd_iommu_v2, amdkfd and radeon
>>>> last. This
>>>> is the order that happens when all three are built as modules. More
>>>> accurately,
>>>> radeon inits, but its init triggers amdkfd init, which triggers
>>>> amd_iommu_v2
>>>> init. So before radeon reaches its probe stage, all the modules were
>>>> initialized.
>>>>
>>>> So what happens when you boot with radeon,
>>>>> amd_iommu_v2 and amdkfd blacklisted for automatically load and
>>>>> only load amdkfd
>>>>> manually?
>>>> As said above, that's ok.
>>>>>
>>>>>> There will always be dependencies between kgd (radeon) and amdkfd
>>>>>> and between
>>>>>> amdkfd and amd_iommu_v2. I don't think I can eliminate those
>>>>>> dependencies, not
>>>>>> without a very complex solution. And the fact that this complex
>>>>>> solution
>>>>>> occurs only in a very specific use case (all modules compiled
>>>>>> in), makes me
>>>>>> less inclined to do that.
>>>>>>
>>>>>> So I don't see it as a "nasty workaround". I would call it just a
>>>>>> "workaround"
>>>>>> for a specific use case, which should be solved by a generic
>>>>>> solution to the
>>>>>> kernel enforcing load orders.
>>>>>
>>>>> The normal kernel module handling already should provide the
>>>>> correct init
>>>>> order,
>>>>> so I would still call this a rather nasty workaround because we
>>>>> couldn't find
>>>>> the underlying problem.
>>>> Well, the normal kernel module handling doesn't work when all
>>>> modules are
>>>> compiled in. I'm not a huge expert on this issue so I had Joerg
>>>> Roedel help me
>>>> analyze this (thanks Joerg) and he agreed that there is no
>>>> enforcement of order
>>>> in this case.
>>>>
>>>>>
>>>>> Christian.
>>>>>
>>>>>>
>>>>>> Oded
>>>>>>>
>>>>>>>> ---
>>>>>>>> drivers/gpu/drm/amd/amdkfd/kfd_module.c | 5 ++---
>>>>>>>> 1 file changed, 2 insertions(+), 3 deletions(-)
>>>>>>>>
>>>>>>>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>>>>>>> b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>>>>>>> index 95d5af1..236562f 100644
>>>>>>>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>>>>>>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>>>>>>> @@ -34,7 +34,7 @@
>>>>>>>> #define KFD_DRIVER_MINOR 7
>>>>>>>> #define KFD_DRIVER_PATCHLEVEL 0
>>>>>>>> -const struct kfd2kgd_calls *kfd2kgd;
>>>>>>>> +const struct kfd2kgd_calls *kfd2kgd = NULL;
>>>>>>>> static const struct kgd2kfd_calls kgd2kfd = {
>>>>>>>> .exit = kgd2kfd_exit,
>>>>>>>> .probe = kgd2kfd_probe,
>>>>>>>> @@ -84,14 +84,13 @@ EXPORT_SYMBOL(kgd2kfd_init);
>>>>>>>> void kgd2kfd_exit(void)
>>>>>>>> {
>>>>>>>> + kfd2kgd = NULL;
>>>>>>>> }
>>>>>>>> static int __init kfd_module_init(void)
>>>>>>>> {
>>>>>>>> int err;
>>>>>>>> - kfd2kgd = NULL;
>>>>>>>> -
>>>>>>>> /* Verify module parameters */
>>>>>>>> if ((sched_policy < KFD_SCHED_POLICY_HWS) ||
>>>>>>>> (sched_policy > KFD_SCHED_POLICY_NO_HWS)) {
>>>>>>>
>>>>>
>>>> _______________________________________________
>>>> dri-devel mailing list
>>>> [email protected]
>>>> http://lists.freedesktop.org/mailman/listinfo/dri-devel
>>
>> --
>> To unsubscribe from this list: send the line "unsubscribe
>> linux-kernel" in
>> the body of a message to [email protected]
>> More majordomo info at http://vger.kernel.org/majordomo-info.html
>> Please read the FAQ at http://www.tux.org/lkml/
On 12/21/2014 06:10 PM, Christian König wrote:
> Am 21.12.2014 um 17:03 schrieb Oded Gabbay:
>>
>>
>> On 12/21/2014 05:57 PM, Christian König wrote:
>>>> There should be, but when the modules are compiled in, they are loaded based on
>>>> link order only, if they are in the same group, and the groups are loaded by a
>>>> pre-defined order.
>>> Is that really still up to date? I've seen effort to change that something like
>>> 10+ years ago when Rusty reworked the module system. And it is comming up on the
>>> lists again from time to time.
>> From what I can see in the Makefile rules, code and google, yes, that's still
>> the situation. If someone will prove me wrong I will be more than happy to
>> correct my code.
>>>
>>>> I don't want to move iommu before gpu, so I don't have a solution for the
>>>> order between amdkfd and amd_iommu_v2.
>>> Why not? That's still better than creating a kernel workqueue, scheduling one
>>> work item on it, rescheduling the task until everything is completed and you can
>>> continue.
>> Because I don't know the consequences of moving an entire subsystem in front
>> of another one. In addition, even if everyone agrees, I'm pretty sure that
>> Linus won't be happy to do that in -rc stages. So maybe this is something to
>> consider for 3.20 merge window, but I would still like to provide a solution
>> for 3.19.
>
> Yeah, true indeed. How about depending on everything being compiled as module
> for 3.19 then? Still better than having such a hack in the driver for as a
> temporary workaround for one release.
>
I thought about it, but because this problem was originally reported by a user
that told us he couldn't use modules because of his setup, I decided not to.
I assume there are other users out there who needs this option (compiled
everything in the kernel - embedded ?), so I don't want to make their life harder.
In addition, saying it is a workaround for one release is true in case moving
iommu subsystem in front of gpu subsystem is acceptable and doesn't cause other
problems, unknown at this point.
Bottom line, my personal preference is to help the users _now_ and if a better
fix is found in the future, change the code accordingly.
Oded
> Christian.
>
>>
>> Oded
>>>
>>> Christian.
>>>
>>> Am 21.12.2014 um 14:24 schrieb Oded Gabbay:
>>>>
>>>>
>>>> On 12/21/2014 03:06 PM, Oded Gabbay wrote:
>>>>>
>>>>>
>>>>> On 12/21/2014 02:19 PM, Christian König wrote:
>>>>>> Am 21.12.2014 um 12:34 schrieb Oded Gabbay:
>>>>>>>
>>>>>>>
>>>>>>> On 12/21/2014 01:27 PM, Christian König wrote:
>>>>>>>> Am 20.12.2014 um 21:46 schrieb Oded Gabbay:
>>>>>>>>> When amdkfd and radeon are compiled inside the kernel image (not as
>>>>>>>>> modules),
>>>>>>>>> radeon will load before amdkfd and will set *kfd2kgd to its interface
>>>>>>>>> structure. Therefore, we must not set *kfd2kgd to NULL when amdkfd is
>>>>>>>>> loaded
>>>>>>>>> because it will override radeon's initialization and cause kernel BUG.
>>>>>>>>>
>>>>>>>>> Signed-off-by: Oded Gabbay <[email protected]>
>>>>>>>>
>>>>>>>> You should probably rather fix the dependency between the two modules to
>>>>>>>> get an
>>>>>>>> determined load order instead of doing such nasty workarounds.
>>>>>>>>
>>>>>>>> Christian.
>>>>>>>
>>>>>>> The problem is that when modules are compiled inside the kernel, there is NO
>>>>>>> determined load order and there is no mechanism to enforce that. If there
>>>>>>> is/was such a mechanism, I would of course prefer to use it.
>>>>>>
>>>>>> There should be an determined order based on the symbol use, otherwise
>>>>>> initializing most of the kernel modules wouldn't work as expected. For
>>>>>> example
>>>>>> radeon depends on the drm module must be loaded before radeon is loaded.
>>>>> There should be, but when the modules are compiled in, they are loaded
>>>>> based on
>>>>> link order only, if they are in the same group, and the groups are loaded by a
>>>>> pre-defined order.
>>>>> The groups are: pure, core, postcore, arch, subsys, fs, device (which
>>>>> represents
>>>>> all the modules) and late. See init.h
>>>>>
>>>>> So radeon, amdkfd and amd_iommu_v2 are all in device group, and in the group
>>>>> they are ordered by their link order.
>>>>>
>>>>> Yes, radeon loads after drm, because drm*.o are before radeon*.o in the
>>>>> Makefile. See
>>>>> http://stackoverflow.com/questions/5669647/linux-order-of-statically-linked-module-loading
>>>>>
>>>>>
>>>>>
>>>>
>>>> So I tried moving amdkfd inside the Makefile before radeon, and that made
>>>> amdkfd load before radeon did. This solves my first problem - order between
>>>> amdkfd and radeon. However, amd_iommu_v2 doesn't belong to the drm Makefile,
>>>> and I don't want to move iommu before gpu, so I don't have a solution for the
>>>> order between amdkfd and amd_iommu_v2.
>>>>
>>>> Oded
>>>>>
>>>>>
>>>>>>
>>>>>>>
>>>>>>> Actually, I don't understand why the kernel doesn't enforce the order
>>>>>>> according to the use of exported symbols, like it does with modules.
>>>>>>
>>>>>> Yeah, that's indeed rather strange. There must be something in the amdkfd
>>>>>> code
>>>>>> which broke that somehow.
>>>>> IMO, that's a far-fetched guess. Could you point to something more specific ?
>>>>>
>>>>>>
>>>>>> As far as I understand you the desired init order is radeon and amd_iommu_v2
>>>>>> first and then amdkfd, right?
>>>>> Actually no. The preferred order is amd_iommu_v2, amdkfd and radeon last. This
>>>>> is the order that happens when all three are built as modules. More
>>>>> accurately,
>>>>> radeon inits, but its init triggers amdkfd init, which triggers amd_iommu_v2
>>>>> init. So before radeon reaches its probe stage, all the modules were
>>>>> initialized.
>>>>>
>>>>> So what happens when you boot with radeon,
>>>>>> amd_iommu_v2 and amdkfd blacklisted for automatically load and only load
>>>>>> amdkfd
>>>>>> manually?
>>>>> As said above, that's ok.
>>>>>>
>>>>>>> There will always be dependencies between kgd (radeon) and amdkfd and
>>>>>>> between
>>>>>>> amdkfd and amd_iommu_v2. I don't think I can eliminate those
>>>>>>> dependencies, not
>>>>>>> without a very complex solution. And the fact that this complex solution
>>>>>>> occurs only in a very specific use case (all modules compiled in), makes me
>>>>>>> less inclined to do that.
>>>>>>>
>>>>>>> So I don't see it as a "nasty workaround". I would call it just a
>>>>>>> "workaround"
>>>>>>> for a specific use case, which should be solved by a generic solution to the
>>>>>>> kernel enforcing load orders.
>>>>>>
>>>>>> The normal kernel module handling already should provide the correct init
>>>>>> order,
>>>>>> so I would still call this a rather nasty workaround because we couldn't find
>>>>>> the underlying problem.
>>>>> Well, the normal kernel module handling doesn't work when all modules are
>>>>> compiled in. I'm not a huge expert on this issue so I had Joerg Roedel help me
>>>>> analyze this (thanks Joerg) and he agreed that there is no enforcement of
>>>>> order
>>>>> in this case.
>>>>>
>>>>>>
>>>>>> Christian.
>>>>>>
>>>>>>>
>>>>>>> Oded
>>>>>>>>
>>>>>>>>> ---
>>>>>>>>> drivers/gpu/drm/amd/amdkfd/kfd_module.c | 5 ++---
>>>>>>>>> 1 file changed, 2 insertions(+), 3 deletions(-)
>>>>>>>>>
>>>>>>>>> diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>>>>>>>> b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>>>>>>>> index 95d5af1..236562f 100644
>>>>>>>>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>>>>>>>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_module.c
>>>>>>>>> @@ -34,7 +34,7 @@
>>>>>>>>> #define KFD_DRIVER_MINOR 7
>>>>>>>>> #define KFD_DRIVER_PATCHLEVEL 0
>>>>>>>>> -const struct kfd2kgd_calls *kfd2kgd;
>>>>>>>>> +const struct kfd2kgd_calls *kfd2kgd = NULL;
>>>>>>>>> static const struct kgd2kfd_calls kgd2kfd = {
>>>>>>>>> .exit = kgd2kfd_exit,
>>>>>>>>> .probe = kgd2kfd_probe,
>>>>>>>>> @@ -84,14 +84,13 @@ EXPORT_SYMBOL(kgd2kfd_init);
>>>>>>>>> void kgd2kfd_exit(void)
>>>>>>>>> {
>>>>>>>>> + kfd2kgd = NULL;
>>>>>>>>> }
>>>>>>>>> static int __init kfd_module_init(void)
>>>>>>>>> {
>>>>>>>>> int err;
>>>>>>>>> - kfd2kgd = NULL;
>>>>>>>>> -
>>>>>>>>> /* Verify module parameters */
>>>>>>>>> if ((sched_policy < KFD_SCHED_POLICY_HWS) ||
>>>>>>>>> (sched_policy > KFD_SCHED_POLICY_NO_HWS)) {
>>>>>>>>
>>>>>>
>>>>> _______________________________________________
>>>>> dri-devel mailing list
>>>>> [email protected]
>>>>> http://lists.freedesktop.org/mailman/listinfo/dri-devel
>>>
>>> --
>>> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
>>> the body of a message to [email protected]
>>> More majordomo info at http://vger.kernel.org/majordomo-info.html
>>> Please read the FAQ at http://www.tux.org/lkml/
>
>>>>> There should be, but when the modules are compiled in, they are loaded
>>>>> based on
>>>>> link order only, if they are in the same group, and the groups are
>>>>> loaded by a
>>>>> pre-defined order.
>>>>
>>>> Is that really still up to date? I've seen effort to change that
>>>> something like
>>>> 10+ years ago when Rusty reworked the module system. And it is comming
>>>> up on the
>>>> lists again from time to time.
>>>
>>> From what I can see in the Makefile rules, code and google, yes, that's
>>> still
>>> the situation. If someone will prove me wrong I will be more than happy
>>> to
>>> correct my code.
>>>>
>>>>
>>>>> I don't want to move iommu before gpu, so I don't have a solution for
>>>>> the
>>>>> order between amdkfd and amd_iommu_v2.
>>>>
>>>> Why not? That's still better than creating a kernel workqueue,
>>>> scheduling one
>>>> work item on it, rescheduling the task until everything is completed and
>>>> you can
>>>> continue.
>>>
>>> Because I don't know the consequences of moving an entire subsystem in
>>> front
>>> of another one. In addition, even if everyone agrees, I'm pretty sure
>>> that
>>> Linus won't be happy to do that in -rc stages. So maybe this is something
>>> to
>>> consider for 3.20 merge window, but I would still like to provide a
>>> solution
>>> for 3.19.
>>
>>
>> Yeah, true indeed. How about depending on everything being compiled as
>> module
>> for 3.19 then? Still better than having such a hack in the driver for as a
>> temporary workaround for one release.
>>
> I thought about it, but because this problem was originally reported by a
> user that told us he couldn't use modules because of his setup, I decided
> not to.
> I assume there are other users out there who needs this option (compiled
> everything in the kernel - embedded ?), so I don't want to make their life
> harder.
>
> In addition, saying it is a workaround for one release is true in case
> moving iommu subsystem in front of gpu subsystem is acceptable and doesn't
> cause other problems, unknown at this point.
>
> Bottom line, my personal preference is to help the users _now_ and if a
> better fix is found in the future, change the code accordingly.
My guess is moving the iommu subsystem in front of the GPU would be rational.
It does seem like it would generally have a depend in that order.
Dave.
On 12/22/2014 09:40 AM, Dave Airlie wrote:
>>>>>> There should be, but when the modules are compiled in, they are loaded
>>>>>> based on
>>>>>> link order only, if they are in the same group, and the groups are
>>>>>> loaded by a
>>>>>> pre-defined order.
>>>>>
>>>>> Is that really still up to date? I've seen effort to change that
>>>>> something like
>>>>> 10+ years ago when Rusty reworked the module system. And it is comming
>>>>> up on the
>>>>> lists again from time to time.
>>>>
>>>> From what I can see in the Makefile rules, code and google, yes, that's
>>>> still
>>>> the situation. If someone will prove me wrong I will be more than happy
>>>> to
>>>> correct my code.
>>>>>
>>>>>
>>>>>> I don't want to move iommu before gpu, so I don't have a solution for
>>>>>> the
>>>>>> order between amdkfd and amd_iommu_v2.
>>>>>
>>>>> Why not? That's still better than creating a kernel workqueue,
>>>>> scheduling one
>>>>> work item on it, rescheduling the task until everything is completed and
>>>>> you can
>>>>> continue.
>>>>
>>>> Because I don't know the consequences of moving an entire subsystem in
>>>> front
>>>> of another one. In addition, even if everyone agrees, I'm pretty sure
>>>> that
>>>> Linus won't be happy to do that in -rc stages. So maybe this is something
>>>> to
>>>> consider for 3.20 merge window, but I would still like to provide a
>>>> solution
>>>> for 3.19.
>>>
>>>
>>> Yeah, true indeed. How about depending on everything being compiled as
>>> module
>>> for 3.19 then? Still better than having such a hack in the driver for as a
>>> temporary workaround for one release.
>>>
>> I thought about it, but because this problem was originally reported by a
>> user that told us he couldn't use modules because of his setup, I decided
>> not to.
>> I assume there are other users out there who needs this option (compiled
>> everything in the kernel - embedded ?), so I don't want to make their life
>> harder.
>>
>> In addition, saying it is a workaround for one release is true in case
>> moving iommu subsystem in front of gpu subsystem is acceptable and doesn't
>> cause other problems, unknown at this point.
>>
>> Bottom line, my personal preference is to help the users _now_ and if a
>> better fix is found in the future, change the code accordingly.
>
> My guess is moving the iommu subsystem in front of the GPU would be rational.
>
> It does seem like it would generally have a depend in that order.
>
> Dave.
>
Dave,
I agree, but don't you think it is too risky for -rc stages ?
If not, I can try it and if it works on KV, I can submit a patch.
But if you do think it is risky, what do you recommend for 3.19 ? Do the fix I
suggested or disable build-in compilation option ?
Oded
Am 22.12.2014 um 08:43 schrieb Oded Gabbay:
>
>
> On 12/22/2014 09:40 AM, Dave Airlie wrote:
>>>>>>> There should be, but when the modules are compiled in, they are
>>>>>>> loaded
>>>>>>> based on
>>>>>>> link order only, if they are in the same group, and the groups are
>>>>>>> loaded by a
>>>>>>> pre-defined order.
>>>>>>
>>>>>> Is that really still up to date? I've seen effort to change that
>>>>>> something like
>>>>>> 10+ years ago when Rusty reworked the module system. And it is
>>>>>> comming
>>>>>> up on the
>>>>>> lists again from time to time.
>>>>>
>>>>> From what I can see in the Makefile rules, code and google, yes,
>>>>> that's
>>>>> still
>>>>> the situation. If someone will prove me wrong I will be more than
>>>>> happy
>>>>> to
>>>>> correct my code.
>>>>>>
>>>>>>
>>>>>>> I don't want to move iommu before gpu, so I don't have a
>>>>>>> solution for
>>>>>>> the
>>>>>>> order between amdkfd and amd_iommu_v2.
>>>>>>
>>>>>> Why not? That's still better than creating a kernel workqueue,
>>>>>> scheduling one
>>>>>> work item on it, rescheduling the task until everything is
>>>>>> completed and
>>>>>> you can
>>>>>> continue.
>>>>>
>>>>> Because I don't know the consequences of moving an entire
>>>>> subsystem in
>>>>> front
>>>>> of another one. In addition, even if everyone agrees, I'm pretty sure
>>>>> that
>>>>> Linus won't be happy to do that in -rc stages. So maybe this is
>>>>> something
>>>>> to
>>>>> consider for 3.20 merge window, but I would still like to provide a
>>>>> solution
>>>>> for 3.19.
>>>>
>>>>
>>>> Yeah, true indeed. How about depending on everything being compiled as
>>>> module
>>>> for 3.19 then? Still better than having such a hack in the driver
>>>> for as a
>>>> temporary workaround for one release.
>>>>
>>> I thought about it, but because this problem was originally reported
>>> by a
>>> user that told us he couldn't use modules because of his setup, I
>>> decided
>>> not to.
>>> I assume there are other users out there who needs this option
>>> (compiled
>>> everything in the kernel - embedded ?), so I don't want to make
>>> their life
>>> harder.
>>>
>>> In addition, saying it is a workaround for one release is true in case
>>> moving iommu subsystem in front of gpu subsystem is acceptable and
>>> doesn't
>>> cause other problems, unknown at this point.
>>>
>>> Bottom line, my personal preference is to help the users _now_ and if a
>>> better fix is found in the future, change the code accordingly.
>>
>> My guess is moving the iommu subsystem in front of the GPU would be
>> rational.
>>
>> It does seem like it would generally have a depend in that order.
>>
>> Dave.
>>
> Dave,
> I agree, but don't you think it is too risky for -rc stages ?
> If not, I can try it and if it works on KV, I can submit a patch.
> But if you do think it is risky, what do you recommend for 3.19 ? Do
> the fix I suggested or disable build-in compilation option ?
I would say create the patch of changing the order (should be trivial),
describe in detail in the commit message what this is supposed to fix
and why such an severe change was done in -rc1 and submit it upstream.
We can still revert it in -rc2 if it breaks anything.
Christian.
>
> Oded
On 12/22/2014 10:57 AM, Christian König wrote:
> Am 22.12.2014 um 08:43 schrieb Oded Gabbay:
>>
>>
>> On 12/22/2014 09:40 AM, Dave Airlie wrote:
>>>>>>>> There should be, but when the modules are compiled in, they are loaded
>>>>>>>> based on
>>>>>>>> link order only, if they are in the same group, and the groups are
>>>>>>>> loaded by a
>>>>>>>> pre-defined order.
>>>>>>>
>>>>>>> Is that really still up to date? I've seen effort to change that
>>>>>>> something like
>>>>>>> 10+ years ago when Rusty reworked the module system. And it is comming
>>>>>>> up on the
>>>>>>> lists again from time to time.
>>>>>>
>>>>>> From what I can see in the Makefile rules, code and google, yes, that's
>>>>>> still
>>>>>> the situation. If someone will prove me wrong I will be more than happy
>>>>>> to
>>>>>> correct my code.
>>>>>>>
>>>>>>>
>>>>>>>> I don't want to move iommu before gpu, so I don't have a solution for
>>>>>>>> the
>>>>>>>> order between amdkfd and amd_iommu_v2.
>>>>>>>
>>>>>>> Why not? That's still better than creating a kernel workqueue,
>>>>>>> scheduling one
>>>>>>> work item on it, rescheduling the task until everything is completed and
>>>>>>> you can
>>>>>>> continue.
>>>>>>
>>>>>> Because I don't know the consequences of moving an entire subsystem in
>>>>>> front
>>>>>> of another one. In addition, even if everyone agrees, I'm pretty sure
>>>>>> that
>>>>>> Linus won't be happy to do that in -rc stages. So maybe this is something
>>>>>> to
>>>>>> consider for 3.20 merge window, but I would still like to provide a
>>>>>> solution
>>>>>> for 3.19.
>>>>>
>>>>>
>>>>> Yeah, true indeed. How about depending on everything being compiled as
>>>>> module
>>>>> for 3.19 then? Still better than having such a hack in the driver for as a
>>>>> temporary workaround for one release.
>>>>>
>>>> I thought about it, but because this problem was originally reported by a
>>>> user that told us he couldn't use modules because of his setup, I decided
>>>> not to.
>>>> I assume there are other users out there who needs this option (compiled
>>>> everything in the kernel - embedded ?), so I don't want to make their life
>>>> harder.
>>>>
>>>> In addition, saying it is a workaround for one release is true in case
>>>> moving iommu subsystem in front of gpu subsystem is acceptable and doesn't
>>>> cause other problems, unknown at this point.
>>>>
>>>> Bottom line, my personal preference is to help the users _now_ and if a
>>>> better fix is found in the future, change the code accordingly.
>>>
>>> My guess is moving the iommu subsystem in front of the GPU would be rational.
>>>
>>> It does seem like it would generally have a depend in that order.
>>>
>>> Dave.
>>>
>> Dave,
>> I agree, but don't you think it is too risky for -rc stages ?
>> If not, I can try it and if it works on KV, I can submit a patch.
>> But if you do think it is risky, what do you recommend for 3.19 ? Do the fix I
>> suggested or disable build-in compilation option ?
>
> I would say create the patch of changing the order (should be trivial), describe
> in detail in the commit message what this is supposed to fix and why such an
> severe change was done in -rc1 and submit it upstream.
>
> We can still revert it in -rc2 if it breaks anything.
>
> Christian.
>
>>
>> Oded
>
OK, I'll try it on my machine and if it works, I will send the patch to the list.
Oded
On 12/22/2014 11:26 AM, Oded Gabbay wrote:
>
>
> On 12/22/2014 10:57 AM, Christian König wrote:
>> Am 22.12.2014 um 08:43 schrieb Oded Gabbay:
>>>
>>>
>>> On 12/22/2014 09:40 AM, Dave Airlie wrote:
>>>>>>>>> There should be, but when the modules are compiled in, they are loaded
>>>>>>>>> based on
>>>>>>>>> link order only, if they are in the same group, and the groups are
>>>>>>>>> loaded by a
>>>>>>>>> pre-defined order.
>>>>>>>>
>>>>>>>> Is that really still up to date? I've seen effort to change that
>>>>>>>> something like
>>>>>>>> 10+ years ago when Rusty reworked the module system. And it is comming
>>>>>>>> up on the
>>>>>>>> lists again from time to time.
>>>>>>>
>>>>>>> From what I can see in the Makefile rules, code and google, yes, that's
>>>>>>> still
>>>>>>> the situation. If someone will prove me wrong I will be more than happy
>>>>>>> to
>>>>>>> correct my code.
>>>>>>>>
>>>>>>>>
>>>>>>>>> I don't want to move iommu before gpu, so I don't have a solution for
>>>>>>>>> the
>>>>>>>>> order between amdkfd and amd_iommu_v2.
>>>>>>>>
>>>>>>>> Why not? That's still better than creating a kernel workqueue,
>>>>>>>> scheduling one
>>>>>>>> work item on it, rescheduling the task until everything is completed and
>>>>>>>> you can
>>>>>>>> continue.
>>>>>>>
>>>>>>> Because I don't know the consequences of moving an entire subsystem in
>>>>>>> front
>>>>>>> of another one. In addition, even if everyone agrees, I'm pretty sure
>>>>>>> that
>>>>>>> Linus won't be happy to do that in -rc stages. So maybe this is something
>>>>>>> to
>>>>>>> consider for 3.20 merge window, but I would still like to provide a
>>>>>>> solution
>>>>>>> for 3.19.
>>>>>>
>>>>>>
>>>>>> Yeah, true indeed. How about depending on everything being compiled as
>>>>>> module
>>>>>> for 3.19 then? Still better than having such a hack in the driver for as a
>>>>>> temporary workaround for one release.
>>>>>>
>>>>> I thought about it, but because this problem was originally reported by a
>>>>> user that told us he couldn't use modules because of his setup, I decided
>>>>> not to.
>>>>> I assume there are other users out there who needs this option (compiled
>>>>> everything in the kernel - embedded ?), so I don't want to make their life
>>>>> harder.
>>>>>
>>>>> In addition, saying it is a workaround for one release is true in case
>>>>> moving iommu subsystem in front of gpu subsystem is acceptable and doesn't
>>>>> cause other problems, unknown at this point.
>>>>>
>>>>> Bottom line, my personal preference is to help the users _now_ and if a
>>>>> better fix is found in the future, change the code accordingly.
>>>>
>>>> My guess is moving the iommu subsystem in front of the GPU would be rational.
>>>>
>>>> It does seem like it would generally have a depend in that order.
>>>>
>>>> Dave.
>>>>
>>> Dave,
>>> I agree, but don't you think it is too risky for -rc stages ?
>>> If not, I can try it and if it works on KV, I can submit a patch.
>>> But if you do think it is risky, what do you recommend for 3.19 ? Do the fix I
>>> suggested or disable build-in compilation option ?
>>
>> I would say create the patch of changing the order (should be trivial), describe
>> in detail in the commit message what this is supposed to fix and why such an
>> severe change was done in -rc1 and submit it upstream.
>>
>> We can still revert it in -rc2 if it breaks anything.
>>
>> Christian.
>>
>>>
>>> Oded
>>
>
> OK, I'll try it on my machine and if it works, I will send the patch to the list.
>
> Oded
> _______________________________________________
> dri-devel mailing list
> [email protected]
> http://lists.freedesktop.org/mailman/listinfo/dri-devel
So I checked it and all my HSA tests are passing on KV machine.
I will send the patches today. Please discard the current patch-set.
Oded