1 Supporting PMUs on RISC-V platforms
2 ==========================================
3 Alan Kao <alankao@andestech.com>, Mar 2018
8 As of this writing, perf_event-related features mentioned in The RISC-V ISA
9 Privileged Version 1.10 are as follows:
10 (please check the manual for more details)
14 * minstret[h], instret[h]
15 * mhpeventx, mhpcounterx[h]
17 With such function set only, porting perf would require a lot of work, due to
18 the lack of the following general architectural performance monitoring features:
20 * Enabling/Disabling counters
21 Counters are just free-running all the time in our case.
22 * Interrupt caused by counter overflow
23 No such feature in the spec.
25 It is not possible to have many interrupt ports for all counters, so an
26 interrupt indicator is required for software to tell which counter has
29 There will be an SBI to support this since the kernel cannot modify the
30 counters [1]. Alternatively, some vendor considers to implement
31 hardware-extension for M-S-U model machines to write counters directly.
33 This document aims to provide developers a quick guide on supporting their
34 PMUs in the kernel. The following sections briefly explain perf' mechanism
37 You may check previous discussions here [1][2]. Also, it might be helpful
38 to check the appendix for related kernel structures.
44 *riscv_pmu* is a global pointer of type *struct riscv_pmu*, which contains
45 various methods according to perf's internal convention and PMU-specific
46 parameters. One should declare such instance to represent the PMU. By default,
47 *riscv_pmu* points to a constant structure *riscv_base_pmu*, which has very
48 basic support to a baseline QEMU model.
50 Then he/she can either assign the instance's pointer to *riscv_pmu* so that
51 the minimal and already-implemented logic can be leveraged, or invent his/her
52 own *riscv_init_platform_pmu* implementation.
54 In other words, existing sources of *riscv_base_pmu* merely provide a
55 reference implementation. Developers can flexibly decide how many parts they
56 can leverage, and in the most extreme case, they can customize every function
57 according to their needs.
60 2. Event Initialization
61 -----------------------
63 When a user launches a perf command to monitor some events, it is first
64 interpreted by the userspace perf tool into multiple *perf_event_open*
65 system calls, and then each of them calls to the body of *event_init*
66 member function that was assigned in the previous step. In *riscv_base_pmu*'s
67 case, it is *riscv_event_init*.
69 The main purpose of this function is to translate the event provided by user
70 into bitmap, so that HW-related control registers or counters can directly be
71 manipulated. The translation is based on the mappings and methods provided in
74 Note that some features can be done in this stage as well:
76 (1) interrupt setting, which is stated in the next section;
77 (2) privilege level setting (user space only, kernel space only, both);
78 (3) destructor setting. Normally it is sufficient to apply *riscv_destroy_event*;
79 (4) tweaks for non-sampling events, which will be utilized by functions such as
80 *perf_adjust_period*, usually something like the follows:
82 if (!is_sampling_event(event)) {
83 hwc->sample_period = x86_pmu.max_period;
84 hwc->last_period = hwc->sample_period;
85 local64_set(&hwc->period_left, hwc->sample_period);
88 In the case of *riscv_base_pmu*, only (3) is provided for now.
94 3.1. Interrupt Initialization
96 This often occurs at the beginning of the *event_init* method. In common
97 practice, this should be a code segment like
99 int x86_reserve_hardware(void)
103 if (!atomic_inc_not_zero(&pmc_refcount)) {
104 mutex_lock(&pmc_reserve_mutex);
105 if (atomic_read(&pmc_refcount) == 0) {
106 if (!reserve_pmc_hardware())
109 reserve_ds_buffers();
112 atomic_inc(&pmc_refcount);
113 mutex_unlock(&pmc_reserve_mutex);
119 And the magic is in *reserve_pmc_hardware*, which usually does atomic
120 operations to make implemented IRQ accessible from some global function pointer.
121 *release_pmc_hardware* serves the opposite purpose, and it is used in event
122 destructors mentioned in previous section.
124 (Note: From the implementations in all the architectures, the *reserve/release*
125 pair are always IRQ settings, so the *pmc_hardware* seems somehow misleading.
126 It does NOT deal with the binding between an event and a physical counter,
127 which will be introduced in the next section.)
131 Basically, a IRQ runs the following pseudo code:
133 for each hardware counter that triggered this overflow
135 get the event of this counter
137 // following two steps are defined as *read()*,
138 // check the section Reading/Writing Counters for details.
139 count the delta value since previous interrupt
140 update the event->count (# event occurs) by adding delta, and
141 event->hw.period_left by subtracting delta
143 if the event overflows
145 set the counter appropriately for the next overflow
147 if the event overflows again
148 too frequently, throttle this event
154 However as of this writing, none of the RISC-V implementations have designed an
155 interrupt for perf, so the details are to be completed in the future.
157 4. Reading/Writing Counters
158 ---------------------------
160 They seem symmetric but perf treats them quite differently. For reading, there
161 is a *read* interface in *struct pmu*, but it serves more than just reading.
162 According to the context, the *read* function not only reads the content of the
163 counter (event->count), but also updates the left period to the next interrupt
164 (event->hw.period_left).
166 But the core of perf does not need direct write to counters. Writing counters
167 is hidden behind the abstraction of 1) *pmu->start*, literally start counting so one
168 has to set the counter to a good value for the next interrupt; 2) inside the IRQ
169 it should set the counter to the same resonable value.
171 Reading is not a problem in RISC-V but writing would need some effort, since
172 counters are not allowed to be written by S-mode.
175 5. add()/del()/start()/stop()
176 -----------------------------
178 Basic idea: add()/del() adds/deletes events to/from a PMU, and start()/stop()
179 starts/stop the counter of some event in the PMU. All of them take the same
180 arguments: *struct perf_event *event* and *int flag*.
182 Consider perf as a state machine, then you will find that these functions serve
183 as the state transition process between those states.
184 Three states (event->hw.state) are defined:
186 * PERF_HES_STOPPED: the counter is stopped
187 * PERF_HES_UPTODATE: the event->count is up-to-date
188 * PERF_HES_ARCH: arch-dependent usage ... we don't need this for now
190 A normal flow of these state transitions are as follows:
192 * A user launches a perf event, resulting in calling to *event_init*.
193 * When being context-switched in, *add* is called by the perf core, with a flag
194 PERF_EF_START, which means that the event should be started after it is added.
195 At this stage, a general event is bound to a physical counter, if any.
196 The state changes to PERF_HES_STOPPED and PERF_HES_UPTODATE, because it is now
197 stopped, and the (software) event count does not need updating.
198 ** *start* is then called, and the counter is enabled.
199 With flag PERF_EF_RELOAD, it writes an appropriate value to the counter (check
200 previous section for detail).
201 Nothing is written if the flag does not contain PERF_EF_RELOAD.
202 The state now is reset to none, because it is neither stopped nor updated
203 (the counting already started)
204 * When being context-switched out, *del* is called. It then checks out all the
205 events in the PMU and calls *stop* to update their counts.
206 ** *stop* is called by *del*
207 and the perf core with flag PERF_EF_UPDATE, and it often shares the same
208 subroutine as *read* with the same logic.
209 The state changes to PERF_HES_STOPPED and PERF_HES_UPTODATE, again.
211 ** Life cycle of these two pairs: *add* and *del* are called repeatedly as
212 tasks switch in-and-out; *start* and *stop* is also called when the perf core
213 needs a quick stop-and-start, for instance, when the interrupt period is being
216 Current implementation is sufficient for now and can be easily extended to
217 features in the future.
219 A. Related Structures
220 ---------------------
222 * struct pmu: include/linux/perf_event.h
223 * struct riscv_pmu: arch/riscv/include/asm/perf_event.h
225 Both structures are designed to be read-only.
227 *struct pmu* defines some function pointer interfaces, and most of them take
228 *struct perf_event* as a main argument, dealing with perf events according to
229 perf's internal state machine (check kernel/events/core.c for details).
231 *struct riscv_pmu* defines PMU-specific parameters. The naming follows the
232 convention of all other architectures.
234 * struct perf_event: include/linux/perf_event.h
235 * struct hw_perf_event
237 The generic structure that represents perf events, and the hardware-related
240 * struct riscv_hw_events: arch/riscv/include/asm/perf_event.h
242 The structure that holds the status of events, has two fixed members:
243 the number of events and the array of the events.
248 [1] https://github.com/riscv/riscv-linux/pull/124
249 [2] https://groups.google.com/a/groups.riscv.org/forum/#!topic/sw-dev/f19TmCNP6yA