1		The MSI Driver Guide HOWTO
2	Tom L Nguyen tom.l.nguyen@intel.com
3			10/03/2003
4	Revised Feb 12, 2004 by Martine Silbermann
5		email: Martine.Silbermann@hp.com
6	Revised Jun 25, 2004 by Tom L Nguyen
7	Revised Jul  9, 2008 by Matthew Wilcox <willy@linux.intel.com>
8		Copyright 2003, 2008 Intel Corporation
9
101. About this guide
11
12This guide describes the basics of Message Signaled Interrupts (MSIs),
13the advantages of using MSI over traditional interrupt mechanisms, how
14to change your driver to use MSI or MSI-X and some basic diagnostics to
15try if a device doesn't support MSIs.
16
17
182. What are MSIs?
19
20A Message Signaled Interrupt is a write from the device to a special
21address which causes an interrupt to be received by the CPU.
22
23The MSI capability was first specified in PCI 2.2 and was later enhanced
24in PCI 3.0 to allow each interrupt to be masked individually.  The MSI-X
25capability was also introduced with PCI 3.0.  It supports more interrupts
26per device than MSI and allows interrupts to be independently configured.
27
28Devices may support both MSI and MSI-X, but only one can be enabled at
29a time.
30
31
323. Why use MSIs?
33
34There are three reasons why using MSIs can give an advantage over
35traditional pin-based interrupts.
36
37Pin-based PCI interrupts are often shared amongst several devices.
38To support this, the kernel must call each interrupt handler associated
39with an interrupt, which leads to reduced performance for the system as
40a whole.  MSIs are never shared, so this problem cannot arise.
41
42When a device writes data to memory, then raises a pin-based interrupt,
43it is possible that the interrupt may arrive before all the data has
44arrived in memory (this becomes more likely with devices behind PCI-PCI
45bridges).  In order to ensure that all the data has arrived in memory,
46the interrupt handler must read a register on the device which raised
47the interrupt.  PCI transaction ordering rules require that all the data
48arrive in memory before the value may be returned from the register.
49Using MSIs avoids this problem as the interrupt-generating write cannot
50pass the data writes, so by the time the interrupt is raised, the driver
51knows that all the data has arrived in memory.
52
53PCI devices can only support a single pin-based interrupt per function.
54Often drivers have to query the device to find out what event has
55occurred, slowing down interrupt handling for the common case.  With
56MSIs, a device can support more interrupts, allowing each interrupt
57to be specialised to a different purpose.  One possible design gives
58infrequent conditions (such as errors) their own interrupt which allows
59the driver to handle the normal interrupt handling path more efficiently.
60Other possible designs include giving one interrupt to each packet queue
61in a network card or each port in a storage controller.
62
63
644. How to use MSIs
65
66PCI devices are initialised to use pin-based interrupts.  The device
67driver has to set up the device to use MSI or MSI-X.  Not all machines
68support MSIs correctly, and for those machines, the APIs described below
69will simply fail and the device will continue to use pin-based interrupts.
70
714.1 Include kernel support for MSIs
72
73To support MSI or MSI-X, the kernel must be built with the CONFIG_PCI_MSI
74option enabled.  This option is only available on some architectures,
75and it may depend on some other options also being set.  For example,
76on x86, you must also enable X86_UP_APIC or SMP in order to see the
77CONFIG_PCI_MSI option.
78
794.2 Using MSI
80
81Most of the hard work is done for the driver in the PCI layer.  It simply
82has to request that the PCI layer set up the MSI capability for this
83device.
84
854.2.1 pci_enable_msi
86
87int pci_enable_msi(struct pci_dev *dev)
88
89A successful call allocates ONE interrupt to the device, regardless
90of how many MSIs the device supports.  The device is switched from
91pin-based interrupt mode to MSI mode.  The dev->irq number is changed
92to a new number which represents the message signaled interrupt;
93consequently, this function should be called before the driver calls
94request_irq(), because an MSI is delivered via a vector that is
95different from the vector of a pin-based interrupt.
96
974.2.2 pci_enable_msi_range
98
99int pci_enable_msi_range(struct pci_dev *dev, int minvec, int maxvec)
100
101This function allows a device driver to request any number of MSI
102interrupts within specified range from 'minvec' to 'maxvec'.
103
104If this function returns a positive number it indicates the number of
105MSI interrupts that have been successfully allocated.  In this case
106the device is switched from pin-based interrupt mode to MSI mode and
107updates dev->irq to be the lowest of the new interrupts assigned to it.
108The other interrupts assigned to the device are in the range dev->irq
109to dev->irq + returned value - 1.  Device driver can use the returned
110number of successfully allocated MSI interrupts to further allocate
111and initialize device resources.
112
113If this function returns a negative number, it indicates an error and
114the driver should not attempt to request any more MSI interrupts for
115this device.
116
117This function should be called before the driver calls request_irq(),
118because MSI interrupts are delivered via vectors that are different
119from the vector of a pin-based interrupt.
120
121It is ideal if drivers can cope with a variable number of MSI interrupts;
122there are many reasons why the platform may not be able to provide the
123exact number that a driver asks for.
124
125There could be devices that can not operate with just any number of MSI
126interrupts within a range.  See chapter 4.3.1.3 to get the idea how to
127handle such devices for MSI-X - the same logic applies to MSI.
128
1294.2.1.1 Maximum possible number of MSI interrupts
130
131The typical usage of MSI interrupts is to allocate as many vectors as
132possible, likely up to the limit returned by pci_msi_vec_count() function:
133
134static int foo_driver_enable_msi(struct pci_dev *pdev, int nvec)
135{
136	return pci_enable_msi_range(pdev, 1, nvec);
137}
138
139Note the value of 'minvec' parameter is 1.  As 'minvec' is inclusive,
140the value of 0 would be meaningless and could result in error.
141
142Some devices have a minimal limit on number of MSI interrupts.
143In this case the function could look like this:
144
145static int foo_driver_enable_msi(struct pci_dev *pdev, int nvec)
146{
147	return pci_enable_msi_range(pdev, FOO_DRIVER_MINIMUM_NVEC, nvec);
148}
149
1504.2.1.2 Exact number of MSI interrupts
151
152If a driver is unable or unwilling to deal with a variable number of MSI
153interrupts it could request a particular number of interrupts by passing
154that number to pci_enable_msi_range() function as both 'minvec' and 'maxvec'
155parameters:
156
157static int foo_driver_enable_msi(struct pci_dev *pdev, int nvec)
158{
159	return pci_enable_msi_range(pdev, nvec, nvec);
160}
161
162Note, unlike pci_enable_msi_exact() function, which could be also used to
163enable a particular number of MSI-X interrupts, pci_enable_msi_range()
164returns either a negative errno or 'nvec' (not negative errno or 0 - as
165pci_enable_msi_exact() does).
166
1674.2.1.3 Single MSI mode
168
169The most notorious example of the request type described above is
170enabling the single MSI mode for a device.  It could be done by passing
171two 1s as 'minvec' and 'maxvec':
172
173static int foo_driver_enable_single_msi(struct pci_dev *pdev)
174{
175	return pci_enable_msi_range(pdev, 1, 1);
176}
177
178Note, unlike pci_enable_msi() function, which could be also used to
179enable the single MSI mode, pci_enable_msi_range() returns either a
180negative errno or 1 (not negative errno or 0 - as pci_enable_msi()
181does).
182
1834.2.3 pci_enable_msi_exact
184
185int pci_enable_msi_exact(struct pci_dev *dev, int nvec)
186
187This variation on pci_enable_msi_range() call allows a device driver to
188request exactly 'nvec' MSIs.
189
190If this function returns a negative number, it indicates an error and
191the driver should not attempt to request any more MSI interrupts for
192this device.
193
194By contrast with pci_enable_msi_range() function, pci_enable_msi_exact()
195returns zero in case of success, which indicates MSI interrupts have been
196successfully allocated.
197
1984.2.4 pci_disable_msi
199
200void pci_disable_msi(struct pci_dev *dev)
201
202This function should be used to undo the effect of pci_enable_msi_range().
203Calling it restores dev->irq to the pin-based interrupt number and frees
204the previously allocated MSIs.  The interrupts may subsequently be assigned
205to another device, so drivers should not cache the value of dev->irq.
206
207Before calling this function, a device driver must always call free_irq()
208on any interrupt for which it previously called request_irq().
209Failure to do so results in a BUG_ON(), leaving the device with
210MSI enabled and thus leaking its vector.
211
2124.2.4 pci_msi_vec_count
213
214int pci_msi_vec_count(struct pci_dev *dev)
215
216This function could be used to retrieve the number of MSI vectors the
217device requested (via the Multiple Message Capable register). The MSI
218specification only allows the returned value to be a power of two,
219up to a maximum of 2^5 (32).
220
221If this function returns a negative number, it indicates the device is
222not capable of sending MSIs.
223
224If this function returns a positive number, it indicates the maximum
225number of MSI interrupt vectors that could be allocated.
226
2274.3 Using MSI-X
228
229The MSI-X capability is much more flexible than the MSI capability.
230It supports up to 2048 interrupts, each of which can be controlled
231independently.  To support this flexibility, drivers must use an array of
232`struct msix_entry':
233
234struct msix_entry {
235	u16 	vector; /* kernel uses to write alloc vector */
236	u16	entry; /* driver uses to specify entry */
237};
238
239This allows for the device to use these interrupts in a sparse fashion;
240for example, it could use interrupts 3 and 1027 and yet allocate only a
241two-element array.  The driver is expected to fill in the 'entry' value
242in each element of the array to indicate for which entries the kernel
243should assign interrupts; it is invalid to fill in two entries with the
244same number.
245
2464.3.1 pci_enable_msix_range
247
248int pci_enable_msix_range(struct pci_dev *dev, struct msix_entry *entries,
249			  int minvec, int maxvec)
250
251Calling this function asks the PCI subsystem to allocate any number of
252MSI-X interrupts within specified range from 'minvec' to 'maxvec'.
253The 'entries' argument is a pointer to an array of msix_entry structs
254which should be at least 'maxvec' entries in size.
255
256On success, the device is switched into MSI-X mode and the function
257returns the number of MSI-X interrupts that have been successfully
258allocated.  In this case the 'vector' member in entries numbered from
2590 to the returned value - 1 is populated with the interrupt number;
260the driver should then call request_irq() for each 'vector' that it
261decides to use.  The device driver is responsible for keeping track of the
262interrupts assigned to the MSI-X vectors so it can free them again later.
263Device driver can use the returned number of successfully allocated MSI-X
264interrupts to further allocate and initialize device resources.
265
266If this function returns a negative number, it indicates an error and
267the driver should not attempt to allocate any more MSI-X interrupts for
268this device.
269
270This function, in contrast with pci_enable_msi_range(), does not adjust
271dev->irq.  The device will not generate interrupts for this interrupt
272number once MSI-X is enabled.
273
274Device drivers should normally call this function once per device
275during the initialization phase.
276
277It is ideal if drivers can cope with a variable number of MSI-X interrupts;
278there are many reasons why the platform may not be able to provide the
279exact number that a driver asks for.
280
281There could be devices that can not operate with just any number of MSI-X
282interrupts within a range.  E.g., an network adapter might need let's say
283four vectors per each queue it provides.  Therefore, a number of MSI-X
284interrupts allocated should be a multiple of four.  In this case interface
285pci_enable_msix_range() can not be used alone to request MSI-X interrupts
286(since it can allocate any number within the range, without any notion of
287the multiple of four) and the device driver should master a custom logic
288to request the required number of MSI-X interrupts.
289
2904.3.1.1 Maximum possible number of MSI-X interrupts
291
292The typical usage of MSI-X interrupts is to allocate as many vectors as
293possible, likely up to the limit returned by pci_msix_vec_count() function:
294
295static int foo_driver_enable_msix(struct foo_adapter *adapter, int nvec)
296{
297	return pci_enable_msix_range(adapter->pdev, adapter->msix_entries,
298				     1, nvec);
299}
300
301Note the value of 'minvec' parameter is 1.  As 'minvec' is inclusive,
302the value of 0 would be meaningless and could result in error.
303
304Some devices have a minimal limit on number of MSI-X interrupts.
305In this case the function could look like this:
306
307static int foo_driver_enable_msix(struct foo_adapter *adapter, int nvec)
308{
309	return pci_enable_msix_range(adapter->pdev, adapter->msix_entries,
310				     FOO_DRIVER_MINIMUM_NVEC, nvec);
311}
312
3134.3.1.2 Exact number of MSI-X interrupts
314
315If a driver is unable or unwilling to deal with a variable number of MSI-X
316interrupts it could request a particular number of interrupts by passing
317that number to pci_enable_msix_range() function as both 'minvec' and 'maxvec'
318parameters:
319
320static int foo_driver_enable_msix(struct foo_adapter *adapter, int nvec)
321{
322	return pci_enable_msix_range(adapter->pdev, adapter->msix_entries,
323				     nvec, nvec);
324}
325
326Note, unlike pci_enable_msix_exact() function, which could be also used to
327enable a particular number of MSI-X interrupts, pci_enable_msix_range()
328returns either a negative errno or 'nvec' (not negative errno or 0 - as
329pci_enable_msix_exact() does).
330
3314.3.1.3 Specific requirements to the number of MSI-X interrupts
332
333As noted above, there could be devices that can not operate with just any
334number of MSI-X interrupts within a range.  E.g., let's assume a device that
335is only capable sending the number of MSI-X interrupts which is a power of
336two.  A routine that enables MSI-X mode for such device might look like this:
337
338/*
339 * Assume 'minvec' and 'maxvec' are non-zero
340 */
341static int foo_driver_enable_msix(struct foo_adapter *adapter,
342				  int minvec, int maxvec)
343{
344	int rc;
345
346	minvec = roundup_pow_of_two(minvec);
347	maxvec = rounddown_pow_of_two(maxvec);
348
349	if (minvec > maxvec)
350		return -ERANGE;
351
352retry:
353	rc = pci_enable_msix_range(adapter->pdev, adapter->msix_entries,
354				   maxvec, maxvec);
355	/*
356	 * -ENOSPC is the only error code allowed to be analyzed
357	 */
358	if (rc == -ENOSPC) {
359		if (maxvec == 1)
360			return -ENOSPC;
361
362		maxvec /= 2;
363
364		if (minvec > maxvec)
365			return -ENOSPC;
366
367		goto retry;
368	}
369
370	return rc;
371}
372
373Note how pci_enable_msix_range() return value is analyzed for a fallback -
374any error code other than -ENOSPC indicates a fatal error and should not
375be retried.
376
3774.3.2 pci_enable_msix_exact
378
379int pci_enable_msix_exact(struct pci_dev *dev,
380			  struct msix_entry *entries, int nvec)
381
382This variation on pci_enable_msix_range() call allows a device driver to
383request exactly 'nvec' MSI-Xs.
384
385If this function returns a negative number, it indicates an error and
386the driver should not attempt to allocate any more MSI-X interrupts for
387this device.
388
389By contrast with pci_enable_msix_range() function, pci_enable_msix_exact()
390returns zero in case of success, which indicates MSI-X interrupts have been
391successfully allocated.
392
393Another version of a routine that enables MSI-X mode for a device with
394specific requirements described in chapter 4.3.1.3 might look like this:
395
396/*
397 * Assume 'minvec' and 'maxvec' are non-zero
398 */
399static int foo_driver_enable_msix(struct foo_adapter *adapter,
400				  int minvec, int maxvec)
401{
402	int rc;
403
404	minvec = roundup_pow_of_two(minvec);
405	maxvec = rounddown_pow_of_two(maxvec);
406
407	if (minvec > maxvec)
408		return -ERANGE;
409
410retry:
411	rc = pci_enable_msix_exact(adapter->pdev,
412				   adapter->msix_entries, maxvec);
413
414	/*
415	 * -ENOSPC is the only error code allowed to be analyzed
416	 */
417	if (rc == -ENOSPC) {
418		if (maxvec == 1)
419			return -ENOSPC;
420
421		maxvec /= 2;
422
423		if (minvec > maxvec)
424			return -ENOSPC;
425
426		goto retry;
427	} else if (rc < 0) {
428		return rc;
429	}
430
431	return maxvec;
432}
433
4344.3.3 pci_disable_msix
435
436void pci_disable_msix(struct pci_dev *dev)
437
438This function should be used to undo the effect of pci_enable_msix_range().
439It frees the previously allocated MSI-X interrupts. The interrupts may
440subsequently be assigned to another device, so drivers should not cache
441the value of the 'vector' elements over a call to pci_disable_msix().
442
443Before calling this function, a device driver must always call free_irq()
444on any interrupt for which it previously called request_irq().
445Failure to do so results in a BUG_ON(), leaving the device with
446MSI-X enabled and thus leaking its vector.
447
4484.3.3 The MSI-X Table
449
450The MSI-X capability specifies a BAR and offset within that BAR for the
451MSI-X Table.  This address is mapped by the PCI subsystem, and should not
452be accessed directly by the device driver.  If the driver wishes to
453mask or unmask an interrupt, it should call disable_irq() / enable_irq().
454
4554.3.4 pci_msix_vec_count
456
457int pci_msix_vec_count(struct pci_dev *dev)
458
459This function could be used to retrieve number of entries in the device
460MSI-X table.
461
462If this function returns a negative number, it indicates the device is
463not capable of sending MSI-Xs.
464
465If this function returns a positive number, it indicates the maximum
466number of MSI-X interrupt vectors that could be allocated.
467
4684.4 Handling devices implementing both MSI and MSI-X capabilities
469
470If a device implements both MSI and MSI-X capabilities, it can
471run in either MSI mode or MSI-X mode, but not both simultaneously.
472This is a requirement of the PCI spec, and it is enforced by the
473PCI layer.  Calling pci_enable_msi_range() when MSI-X is already
474enabled or pci_enable_msix_range() when MSI is already enabled
475results in an error.  If a device driver wishes to switch between MSI
476and MSI-X at runtime, it must first quiesce the device, then switch
477it back to pin-interrupt mode, before calling pci_enable_msi_range()
478or pci_enable_msix_range() and resuming operation.  This is not expected
479to be a common operation but may be useful for debugging or testing
480during development.
481
4824.5 Considerations when using MSIs
483
4844.5.1 Choosing between MSI-X and MSI
485
486If your device supports both MSI-X and MSI capabilities, you should use
487the MSI-X facilities in preference to the MSI facilities.  As mentioned
488above, MSI-X supports any number of interrupts between 1 and 2048.
489In contrast, MSI is restricted to a maximum of 32 interrupts (and
490must be a power of two).  In addition, the MSI interrupt vectors must
491be allocated consecutively, so the system might not be able to allocate
492as many vectors for MSI as it could for MSI-X.  On some platforms, MSI
493interrupts must all be targeted at the same set of CPUs whereas MSI-X
494interrupts can all be targeted at different CPUs.
495
4964.5.2 Spinlocks
497
498Most device drivers have a per-device spinlock which is taken in the
499interrupt handler.  With pin-based interrupts or a single MSI, it is not
500necessary to disable interrupts (Linux guarantees the same interrupt will
501not be re-entered).  If a device uses multiple interrupts, the driver
502must disable interrupts while the lock is held.  If the device sends
503a different interrupt, the driver will deadlock trying to recursively
504acquire the spinlock.  Such deadlocks can be avoided by using
505spin_lock_irqsave() or spin_lock_irq() which disable local interrupts
506and acquire the lock (see Documentation/DocBook/kernel-locking).
507
5084.6 How to tell whether MSI/MSI-X is enabled on a device
509
510Using 'lspci -v' (as root) may show some devices with "MSI", "Message
511Signalled Interrupts" or "MSI-X" capabilities.  Each of these capabilities
512has an 'Enable' flag which is followed with either "+" (enabled)
513or "-" (disabled).
514
515
5165. MSI quirks
517
518Several PCI chipsets or devices are known not to support MSIs.
519The PCI stack provides three ways to disable MSIs:
520
5211. globally
5222. on all devices behind a specific bridge
5233. on a single device
524
5255.1. Disabling MSIs globally
526
527Some host chipsets simply don't support MSIs properly.  If we're
528lucky, the manufacturer knows this and has indicated it in the ACPI
529FADT table.  In this case, Linux automatically disables MSIs.
530Some boards don't include this information in the table and so we have
531to detect them ourselves.  The complete list of these is found near the
532quirk_disable_all_msi() function in drivers/pci/quirks.c.
533
534If you have a board which has problems with MSIs, you can pass pci=nomsi
535on the kernel command line to disable MSIs on all devices.  It would be
536in your best interests to report the problem to linux-pci@vger.kernel.org
537including a full 'lspci -v' so we can add the quirks to the kernel.
538
5395.2. Disabling MSIs below a bridge
540
541Some PCI bridges are not able to route MSIs between busses properly.
542In this case, MSIs must be disabled on all devices behind the bridge.
543
544Some bridges allow you to enable MSIs by changing some bits in their
545PCI configuration space (especially the Hypertransport chipsets such
546as the nVidia nForce and Serverworks HT2000).  As with host chipsets,
547Linux mostly knows about them and automatically enables MSIs if it can.
548If you have a bridge unknown to Linux, you can enable
549MSIs in configuration space using whatever method you know works, then
550enable MSIs on that bridge by doing:
551
552       echo 1 > /sys/bus/pci/devices/$bridge/msi_bus
553
554where $bridge is the PCI address of the bridge you've enabled (eg
5550000:00:0e.0).
556
557To disable MSIs, echo 0 instead of 1.  Changing this value should be
558done with caution as it could break interrupt handling for all devices
559below this bridge.
560
561Again, please notify linux-pci@vger.kernel.org of any bridges that need
562special handling.
563
5645.3. Disabling MSIs on a single device
565
566Some devices are known to have faulty MSI implementations.  Usually this
567is handled in the individual device driver, but occasionally it's necessary
568to handle this with a quirk.  Some drivers have an option to disable use
569of MSI.  While this is a convenient workaround for the driver author,
570it is not good practice, and should not be emulated.
571
5725.4. Finding why MSIs are disabled on a device
573
574From the above three sections, you can see that there are many reasons
575why MSIs may not be enabled for a given device.  Your first step should
576be to examine your dmesg carefully to determine whether MSIs are enabled
577for your machine.  You should also check your .config to be sure you
578have enabled CONFIG_PCI_MSI.
579
580Then, 'lspci -t' gives the list of bridges above a device.  Reading
581/sys/bus/pci/devices/*/msi_bus will tell you whether MSIs are enabled (1)
582or disabled (0).  If 0 is found in any of the msi_bus files belonging
583to bridges between the PCI root and the device, MSIs are disabled.
584
585It is also worth checking the device driver to see whether it supports MSIs.
586For example, it may contain calls to pci_enable_msi_range() or
587pci_enable_msix_range().
588