DOE WARN splat be gone
cover DOE WARN splat be gone
Here's my proposal to fix the DOE WARN splat reported by Gregory Price:
One backportable oneliner to replace INIT_WORK() with INIT_WORK_ONSTACK() and one patch to change the API and thus fix the problem for good.
So that's an alternative approach to Ira's.
Please review and test. Thanks!
Lukas Wunner (2):
PCI/DOE: Silence WARN splat upon task submission PCI/DOE: Provide synchronous API
drivers/cxl/core/pci.c | 58 +++++++------------- drivers/pci/doe.c | 114 +++++++++++++++++++++++++++++++++------- include/linux/pci-doe.h | 47 ++--------------- 3 files changed, 117 insertions(+), 102 deletions(-)
Signature · 1 line
2.36.1
PATCH 1/2 PCI/DOE: Silence WARN splat upon task submission
Gregory Price reports a WARN splat with CONFIG_DEBUG_OBJECTS=y upon CXL probing because pci_doe_submit_task() invokes INIT_WORK() instead of INIT_WORK_ONSTACK() for a work_struct that was allocated on the stack.
All callers of pci_doe_submit_task() allocate the work_struct on the stack, so replace INIT_WORK() with INIT_WORK_ONSTACK() as a backportable short-term fix.
Stacktrace for posterity:
WARNING: CPU: 0 PID: 23 at lib/debugobjects.c:545 __debug_object_init.cold+0x18/0x183 CPU: 0 PID: 23 Comm: kworker/u2:1 Not tainted 6.1.0-0.rc1.20221019gitaae703b02f92.17.fc38.x86_64 #1 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.16.0-0-gd239552ce722-prebuilt.qemu.org 04/01/2014 Call Trace: pci_doe_submit_task+0x5d/0xd0 pci_doe_discovery+0xb4/0x100 pcim_doe_create_mb+0x219/0x290 cxl_pci_probe+0x192/0x430 local_pci_probe+0x41/0x80 pci_device_probe+0xb3/0x220 really_probe+0xde/0x380 __driver_probe_device+0x78/0x170 driver_probe_device+0x1f/0x90 __driver_attach_async_helper+0x5c/0xe0 async_run_entry_fn+0x30/0x130 process_one_work+0x294/0x5b0
--- drivers/pci/doe.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/pci/doe.c +1 -1
@@ -541,7 +541,7 @@ int pci_doe_submit_task(struct pci_doe_mb *doe_mb, struct pci_doe_task *task)return -EIO ; task->doe_mb = doe_mb; -INIT_WORK (&task->work, doe_statemachine_work); +INIT_WORK_ONSTACK (&task->work, doe_statemachine_work);queue_work (doe_mb->work_queue, &task->work);return 0 ; }
Signature · 1 line
2.36.1
Re: [PATCH 1/2] PCI/DOE: Silence WARN splat upon task submission
Quoted 43 lines · On Mon, 28 Nov 2022 05:25:51 +0100 Lukas Wunner <[email protected]> wrote:
Gregory Price reports a WARN splat with
CONFIG_DEBUG_OBJECTS=yupon CXL probing becausepci_doe_submit_task()invokesINIT_WORK()instead ofINIT_WORK_ONSTACK()for awork_structthat was allocated on the stack.All callers of
pci_doe_submit_task()allocate thework_structon the stack, so replaceINIT_WORK()withINIT_WORK_ONSTACK()as a backportable short-term fix.Stacktrace for posterity:
WARNING: CPU: 0 PID: 23 at lib/debugobjects.c:545 __debug_object_init.cold+0x18/0x183 CPU: 0 PID: 23 Comm: kworker/u2:1 Not tainted 6.1.0-0.rc1.20221019gitaae703b02f92.17.fc38.x86_64 #1 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.16.0-0-gd239552ce722-prebuilt.qemu.org 04/01/2014 Call Trace: pci_doe_submit_task+0x5d/0xd0 pci_doe_discovery+0xb4/0x100 pcim_doe_create_mb+0x219/0x290 cxl_pci_probe+0x192/0x430 local_pci_probe+0x41/0x80 pci_device_probe+0xb3/0x220 really_probe+0xde/0x380 __driver_probe_device+0x78/0x170 driver_probe_device+0x1f/0x90 __driver_attach_async_helper+0x5c/0xe0 async_run_entry_fn+0x30/0x130 process_one_work+0x294/0x5b0Fixes 9d24322e887b ("PCI/DOE: Add DOE mailbox support functions")Reported-by Gregory Price <[email protected]>Signed-off-by Lukas Wunner <[email protected]>Cc [email protected] # v6.0+Cc Ira Weiny <[email protected]>Cc Jonathan Cameron <[email protected]>--- drivers/pci/doe.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)drivers/pci/doe.c +0
-0
drivers/pci/doe.c int pci_doe_submit_task(struct pci_doe_mb *doe_mb, struct pci_doe_task *task) · line 544
Quoted 5 lines
drivers/pci/doe.c +1
-1@@ -541,7 +541,7 @@ int pci_doe_submit_task(struct pci_doe_mb *doe_mb, struct pci_doe_task *task)return -EIO ; task->doe_mb = doe_mb; -INIT_WORK (&task->work, doe_statemachine_work); +INIT_WORK_ONSTACK (&task->work, doe_statemachine_work);
If we go this way, add a comment to say 'why' it is ONSTACK or add to the function description to say it 'must be on stack'.
Quoted 3 lines
drivers/pci/doe.c +0
-0queue_work (doe_mb->work_queue, &task->work);return 0 ; }
Re: [PATCH 1/2] PCI/DOE: Silence WARN splat upon task submission
Quoted 48 lines · Jonathan Cameron
Quoted 45 lines · On Mon, 28 Nov 2022 05:25:51 +0100 Lukas Wunner <[email protected]> wrote:
Gregory Price reports a WARN splat with
CONFIG_DEBUG_OBJECTS=yupon CXL probing becausepci_doe_submit_task()invokesINIT_WORK()instead ofINIT_WORK_ONSTACK()for awork_structthat was allocated on the stack.All callers of
pci_doe_submit_task()allocate thework_structon the stack, so replaceINIT_WORK()withINIT_WORK_ONSTACK()as a backportable short-term fix.Stacktrace for posterity:
WARNING: CPU: 0 PID: 23 at lib/debugobjects.c:545 __debug_object_init.cold+0x18/0x183 CPU: 0 PID: 23 Comm: kworker/u2:1 Not tainted 6.1.0-0.rc1.20221019gitaae703b02f92.17.fc38.x86_64 #1 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.16.0-0-gd239552ce722-prebuilt.qemu.org 04/01/2014 Call Trace: pci_doe_submit_task+0x5d/0xd0 pci_doe_discovery+0xb4/0x100 pcim_doe_create_mb+0x219/0x290 cxl_pci_probe+0x192/0x430 local_pci_probe+0x41/0x80 pci_device_probe+0xb3/0x220 really_probe+0xde/0x380 __driver_probe_device+0x78/0x170 driver_probe_device+0x1f/0x90 __driver_attach_async_helper+0x5c/0xe0 async_run_entry_fn+0x30/0x130 process_one_work+0x294/0x5b0Fixes 9d24322e887b ("PCI/DOE: Add DOE mailbox support functions")Reported-by Gregory Price <[email protected]>Signed-off-by Lukas Wunner <[email protected]>Cc [email protected] # v6.0+Cc Ira Weiny <[email protected]>Cc Jonathan Cameron <[email protected]>--- drivers/pci/doe.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)drivers/pci/doe.c +0
-0@@ -541,7 +541,7 @@ int pci_doe_submit_task(struct pci_doe_mb *doe_mb, struct pci_doe_task *task)return -EIO ;
Quoted 6 lines
drivers/pci/doe.c int pci_doe_submit_task(struct pci_doe_mb *doe_mb, struct pci_doe_task *task) · line 544
Quoted 3 lines
drivers/pci/doe.c +1
-1task->doe_mb = doe_mb; -INIT_WORK (&task->work, doe_statemachine_work); +INIT_WORK_ONSTACK (&task->work, doe_statemachine_work);If we go this way, add a comment to say 'why' it is ONSTACK or add to the function description to say it 'must be on stack'.
My apologies, I somehow missed this series with the chaos of my personal life.
I do like this as a backportable fix and I'm not opposed to it.
I think Lukas will need to determine how much he needs the async support. So I'm going to back off and let him deal with this.
Ira
Quoted 5 lines
Quoted 4 lines
drivers/pci/doe.c +0
-0queue_work (doe_mb->work_queue, &task->work);return 0 ; }
PATCH 2/2 PCI/DOE: Provide synchronous API
In reply to Lukas WunnerThe DOE API only allows asynchronous exchanges and forces callers to provide a completion callback. Yet all existing callers only perform synchronous exchanges. Upcoming patches for CMA (Component Measurement and Authentication, PCIe r6.0.1 sec 6.31) likewise require only synchronous DOE exchanges. Asynchronous users are currently not foreseeable.
Provide a synchronous pci_doe() API call which builds on the internal asynchronous machinery. Should asynchronous users appear, reintroducing a pci_doe_async() API call will be trivial.
Convert all users to the new synchronous API and make the asynchronous pci_doe_submit_task() as well as the pci_doe_task struct private.
--- drivers/cxl/core/pci.c | 58 +++++++-------------- drivers/pci/doe.c | 112 ++++++++++++++++++++++++++++++++++------ include/linux/pci-doe.h | 47 ++--------------- 3 files changed, 116 insertions(+), 101 deletions(-)
drivers/cxl/core/pci.c +18 -40
@@ -490,51 +490,26 @@ static struct pci_doe_mb *find_cdat_doe(struct device *uport) CXL_DOE_TABLE_ACCESS_TABLE_TYPE_CDATA) | \ FIELD_PREP(CXL_DOE_TABLE_ACCESS_ENTRY_HANDLE, (entry_handle))) -static void cxl_doe_task_complete (struct pci_doe_task *task) -{ -complete (task->private); -} - -struct cdat_doe_task { -u32 request_pl; -u32 response_pl[32 ]; -struct completion c; -struct pci_doe_task task; -}; - -#define DECLARE_CDAT_DOE_TASK (req, cdt) \ -struct cdat_doe_task cdt = { \ - .c = COMPLETION_INITIALIZER_ONSTACK(cdt.c), \ - .request_pl = req, \ - .task = { \ - .prot.vid = PCI_DVSEC_VENDOR_ID_CXL, \ - .prot.type = CXL_DOE_PROTOCOL_TABLE_ACCESS, \ - .request_pl = &cdt.request_pl, \ - .request_pl_sz = sizeof(cdt.request_pl), \ - .response_pl = cdt.response_pl, \ - .response_pl_sz = sizeof(cdt.response_pl), \ - .complete = cxl_doe_task_complete, \ - .private = &cdt.c, \ - } \ -} -static int cxl_cdat_get_length (struct device *dev,struct pci_doe_mb *cdat_doe,size_t *length) { -DECLARE_CDAT_DOE_TASK (CDAT_DOE_REQ (0 ), t); +u32 request =CDAT_DOE_REQ (0 ); +u32 response[32 ];int rc; - rc =pci_doe_submit_task (cdat_doe, &t.task); + rc =pci_doe (cdat_doe,PCI_DVSEC_VENDOR_ID_CXL , +CXL_DOE_PROTOCOL_TABLE_ACCESS , + &request,sizeof (request), + &response,sizeof (response));if (rc <0 ) {dev_err (dev,"DOE submit failed: %d" , rc);return rc; } -wait_for_completion (&t.c); -if (t.task.rv <sizeof (u32)) +if (rc <sizeof (u32))return -EIO ; - *length = t.response_pl[1 ]; + *length = response[1 ];dev_dbg (dev,"CDAT length %zu\n" , *length);return 0 ;
@@ -549,26 +524,29 @@ static int cxl_cdat_read_table(struct device *dev,int entry_handle =0 ; do { - DECLARE_CDAT_DOE_TASK(CDAT_DOE_REQ(entry_handle), t); + u32 request = CDAT_DOE_REQ(entry_handle); + u32 response[32]; size_t entry_dw; u32 *entry; int rc; - rc = pci_doe_submit_task(cdat_doe, &t.task); + rc = pci_doe(cdat_doe, PCI_DVSEC_VENDOR_ID_CXL, + CXL_DOE_PROTOCOL_TABLE_ACCESS, + &request, sizeof(request), + &response, sizeof(response)); if (rc < 0) { dev_err(dev, "DOE submit failed: %d", rc); return rc; } - wait_for_completion(&t.c); /* 1 DW header + 1 DW data min */ - if (t.task.rv < (2 * sizeof(u32))) + if (rc < (2 * sizeof(u32))) return -EIO; /* Get the CXL table access header entry handle */ entry_handle = FIELD_GET(CXL_DOE_TABLE_ACCESS_ENTRY_HANDLE, - t.response_pl[0]); - entry = t.response_pl + 1; - entry_dw = t.task.rv / sizeof(u32); + response[0]); + entry = response + 1; + entry_dw = rc / sizeof(u32); /* Skip Header */ entry_dw -= 1; entry_dw = min(length / sizeof(u32), entry_dw);
drivers/pci/doe.c +95 -17
138 lines not on this page · raw message
include/linux/pci-doe.h +3 -44
60 lines not on this page · raw message
Signature · 1 line
2.36.1
Re: [PATCH 2/2] PCI/DOE: Provide synchronous API
Quoted 7 lines · On Mon, 28 Nov 2022 05:25:52 +0100 Lukas Wunner <[email protected]> wrote:
The DOE API only allows asynchronous exchanges and forces callers to provide a completion callback. Yet all existing callers only perform synchronous exchanges. Upcoming patches for CMA (Component Measurement and Authentication, PCIe r6.0.1 sec 6.31) likewise require only synchronous DOE exchanges. Asynchronous users are currently not foreseeable.
Quoted 8 lines
Provide a synchronous
pci_doe()API call which builds on the internal asynchronous machinery. Should asynchronous users appear, reintroducing apci_doe_async()API call will be trivial.Convert all users to the new synchronous API and make the asynchronous
pci_doe_submit_task()as well as thepci_doe_taskstruct private.Signed-off-by Lukas Wunner <[email protected]>
Hi Lukas,
Thanks for looking at this. A few trivial comments line.
This covers the existing question around async vs sync but doesn't have the potential advantages that Ira's series has in terms of ripping out a bunch of complexity.
I'm too tied up in the various implementations to offer a clear view on which way was should go on this - I'll end up spending all day arguing with myself!
It's a bit of crystal ball gazing for how useful keeping the async stuff around will be. Might be a case of taking your first patch then sitting on the current implementation for a cycle or two to see if it get users... Or take approach Ira proposed and only put the infrastructure back in when we have a user for async.
Jonathan
Quoted 4 lines
drivers/pci/doe.c +0
-0
...
Quoted 14 lines
+/**
+ * structpci_doe_task- represents a single query/response
+ *
+ * @prot: DOE Protocol
+ * @request_pl: The request payload
+ * @request_pl_sz: Size of the request payload (bytes)
+ * @response_pl: The response payload
+ * @response_pl_sz: Size of the response payload (bytes)
+ * @rv: Return value. Length of received response or error (bytes)
+ * @complete: Called when task is complete
+ * @private: Private data for the consumer
+ * @work: Used internally by the mailbox
+ * @doe_mb: Used internally by the mailbox
+ *
Quoted 8 lines
+ * The payload sizes and rv are specified in bytes with the following
+ * restrictions concerning the protocol.
+ *
+ * 1) Therequest_pl_szmust be a multiple of double words (4 bytes)
+ * 2) Theresponse_pl_szmust be >= a single double word (4 bytes)
+ * 3) rv is returned as bytes but it will be a multiple of double words
+ *
+ * NOTE there is no need for the caller to initialize work ordoe_mb.
Cut and paste from original, but what's the "caller" of a struct? I'd just drop this NOTE as it's better explained below.
Quoted 16 lines
+ */
drivers/pci/doe.c +15
-0+struct pci_doe_task { +struct pci_doe_protocol prot; +u32 *request_pl; +size_t request_pl_sz; +u32 *response_pl; +size_t response_pl_sz; +int rv; +void (*complete)(struct pci_doe_task *task); +void *private; + +/* initialized by pci_doe_submit_task() */ +struct work_struct work; +struct pci_doe_mb *doe_mb; +}; +
...
drivers/pci/doe.c struct pci_doe_task { · line 32
Quoted 9 lines
/**
*pci_doe_for_each_off- Iterate each DOE capability
* @pdev: structpci_devto iteratedrivers/pci/doe.c +1
-1@@ -72,6 +29,8 @@ struct pci_doe_task {struct pci_doe_mb *pcim_doe_create_mb (struct pci_dev *pdev,u16 cap_offset);bool pci_doe_supports_prot (struct pci_doe_mb *doe_mb,u16 vid,u8 type); -int pci_doe_submit_task (struct pci_doe_mb *doe_mb,struct pci_doe_task *task); +int pci_doe(struct pci_doe_mb *doe_mb, u16 vendor, u8 type,
Whilst there is clearly a verb hidden in that doe, the fact that the whole spec section is called the same is confusing.
pci_doe_query_response() maybe or pci_doe_do() perhaps?
Quoted 4 lines
drivers/pci/doe.c +2
-0+ void *request, size_t request_sz, + void *response, size_t response_sz);#endif
Re: [PATCH 2/2] PCI/DOE: Provide synchronous API
Quoted 27 lines · Jonathan Cameron
Quoted 7 lines · On Mon, 28 Nov 2022 05:25:52 +0100 Lukas Wunner <[email protected]> wrote:
The DOE API only allows asynchronous exchanges and forces callers to provide a completion callback. Yet all existing callers only perform synchronous exchanges. Upcoming patches for CMA (Component Measurement and Authentication, PCIe r6.0.1 sec 6.31) likewise require only synchronous DOE exchanges. Asynchronous users are currently not foreseeable.
Quoted 8 lines
Provide a synchronous
pci_doe()API call which builds on the internal asynchronous machinery. Should asynchronous users appear, reintroducing apci_doe_async()API call will be trivial.Convert all users to the new synchronous API and make the asynchronous
pci_doe_submit_task()as well as thepci_doe_taskstruct private.Signed-off-by Lukas Wunner <[email protected]>Hi Lukas,
Thanks for looking at this. A few trivial comments line.
This covers the existing question around async vs sync but doesn't have the potential advantages that Ira's series has in terms of ripping out a bunch of complexity.
Quoted 5 lines
I'm too tied up in the various implementations to offer a clear view on which way was should go on this - I'll end up spending all day arguing with myself!
It's a bit of crystal ball gazing for how useful keeping the async stuff
I agree that this is much too 'crystal ball gazing' for me as well. See below for more.
Quoted 64 lines
around will be. Might be a case of taking your first patch then sitting on the current implementation for a cycle or two to see if it get users... Or take approach Ira proposed and only put the infrastructure back in when we have a user for async.
Jonathan
Quoted 4 lines
drivers/pci/doe.c +0
-0...
Quoted 14 lines
+/**
+ * structpci_doe_task- represents a single query/response
+ *
+ * @prot: DOE Protocol
+ * @request_pl: The request payload
+ * @request_pl_sz: Size of the request payload (bytes)
+ * @response_pl: The response payload
+ * @response_pl_sz: Size of the response payload (bytes)
+ * @rv: Return value. Length of received response or error (bytes)
+ * @complete: Called when task is complete
+ * @private: Private data for the consumer
+ * @work: Used internally by the mailbox
+ * @doe_mb: Used internally by the mailbox
+ *Quoted 8 lines
+ * The payload sizes and rv are specified in bytes with the following
+ * restrictions concerning the protocol.
+ *
+ * 1) Therequest_pl_szmust be a multiple of double words (4 bytes)
+ * 2) Theresponse_pl_szmust be >= a single double word (4 bytes)
+ * 3) rv is returned as bytes but it will be a multiple of double words
+ *
+ * NOTE there is no need for the caller to initialize work ordoe_mb.Cut and paste from original, but what's the "caller" of a struct? I'd just drop this NOTE as it's better explained below.
Quoted 16 lines
+ */
drivers/pci/doe.c +15
-0+struct pci_doe_task { +struct pci_doe_protocol prot; +u32 *request_pl; +size_t request_pl_sz; +u32 *response_pl; +size_t response_pl_sz; +int rv; +void (*complete)(struct pci_doe_task *task); +void *private; + +/* initialized by pci_doe_submit_task() */ +struct work_struct work; +struct pci_doe_mb *doe_mb; +}; +...
Quoted 5 lines
/**
*pci_doe_for_each_off- Iterate each DOE capability
* @pdev: structpci_devto iteratedrivers/pci/doe.c +0
-0@@ -72,6 +29,8 @@ struct pci_doe_task {
Quoted 8 lines
drivers/pci/doe.c struct pci_doe_task { · line 32
Quoted 4 lines
drivers/pci/doe.c +1
-1struct pci_doe_mb *pcim_doe_create_mb (struct pci_dev *pdev,u16 cap_offset);bool pci_doe_supports_prot (struct pci_doe_mb *doe_mb,u16 vid,u8 type); -int pci_doe_submit_task (struct pci_doe_mb *doe_mb,struct pci_doe_task *task); +int pci_doe(struct pci_doe_mb *doe_mb, u16 vendor, u8 type,Whilst there is clearly a verb hidden in that doe, the fact that the whole spec section is called the same is confusing.
pci_doe_query_response()maybe orpci_doe_do()perhaps?
Or just pci_doe_submit()?
Lukas and I discussed this off-line. Because he is going to need this stuff going forward. I'm going to back off fixing this and let him handle it.
I agree with him that eventually something like a 'flush' operation will be needed but right now that mechanism is broken. I'll let him determine if it should be removed or fixed depending on his future needs.
Ira
Quoted 7 lines
Quoted 6 lines
drivers/pci/doe.c +2
-0+ void *request, size_t request_sz, + void *response, size_t response_sz);#endif
Re: [PATCH 2/2] PCI/DOE: Provide synchronous API
In reply to Jonathan CameronQuoted 5 lines · Jonathan Cameron
Quoted 1 line · On Mon, 28 Nov 2022 05:25:52 +0100 Lukas Wunner <[email protected]> wrote:
+ * NOTE there is no need for the caller to initialize work or
doe_mb.Cut and paste from original, but what's the "caller" of a struct? I'd just drop this NOTE as it's better explained below.
Okay, will drop this note when respinning, it just duplicates the code comment inside the pci_doe_task declaration.
Quoted 9 lines
include/linux/pci-doe.h struct pci_doe_task { · line 32
Quoted 4 lines
struct
pci_doe_mb*pcim_doe_create_mb(structpci_dev*pdev, u16cap_offset); boolpci_doe_supports_prot(struct pci_doe_mb *doe_mb, u16 vid, u8 type);include/linux/pci-doe.h +1
-1-int pci_doe_submit_task (struct pci_doe_mb *doe_mb,struct pci_doe_task *task); +int pci_doe(struct pci_doe_mb *doe_mb, u16 vendor, u8 type,Whilst there is clearly a verb hidden in that doe, the fact that the whole spec section is called the same is confusing.
pci_doe_query_response()maybe orpci_doe_do()perhaps?
Going forward the PCI core will allocate DOE mailboxes on device enumeration. And the API will then consist of just two calls:
One call to get a pointer to a pci_doe_mb for a (pci_dev, vendor, protocol) triple which I intend to call pci_find_doe_mailbox().
And one call to perform a synchronous query/response over a pci_doe_mb. That's the pci_doe() I'm introducing here.
I initially went for pci_doe_exchange(), then realized that the "e" in "doe" already stands for "exchange", so that would be redundant.
We generally try to use the same terms as the PCIe Base Spec to make it easy for an uninitiated reader to connect the spec to the implementation. The spec talks about "performing data object exchanges", but pci_doe_perform() didn't seem very appealing to me.
I ended up just using pci_doe() for simplicity and because it makes it easy to stay within 80 chars if the function name is short.
Passing in the request and response pointers directly instead of populating a struct is likewise an attempt to make things as simple as possible for driver authors. It's modeled after the SPI and I2C APIs which offer similar primitives to perform read/write exchanges with devices.
(Side note: pci_find_doe_mailbox() will return the *first* mailbox found for a given (pci_dev, vendor, protocol) triple. Conceivably, there may be devices with multiple mailboxes supporting the same protocols. We may have to add a pci_find_next_doe_mailbox(), should such devices appear. Those names match what we already have for capabilities, i.e. pci_find_ext_capability() and pci_find_next_ext_capability().)
Thanks,
Lukas