[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <ae674391-c8b1-47f9-b92f-748bf376b4a5@arm.com>
Date: Fri, 16 Aug 2024 14:21:52 +0100
From: Robin Murphy <robin.murphy@....com>
To: Mark Rutland <mark.rutland@....com>
Cc: will@...nel.org, linux-arm-kernel@...ts.infradead.org,
linux-kernel@...r.kernel.org, ilkka@...amperecomputing.com
Subject: Re: [PATCH 1/8] perf/arm-cmn: Refactor node ID handling. Again.
On 16/08/2024 1:45 pm, Mark Rutland wrote:
> On Fri, Aug 16, 2024 at 12:56:48PM +0100, Robin Murphy wrote:
>> On 2024-08-16 10:33 am, Mark Rutland wrote:
>>> Hi Robin,
>>>
>>> On Fri, Aug 09, 2024 at 08:15:40PM +0100, Robin Murphy wrote:
>>>> It transpires that despite the explicit example to the contrary in the
>>>> CMN-700 TRM, the "extra device ports" config is in fact a per-XP thing
>>>> rather than a global one. To that end, rework node IDs yet again to
>>>> carry around the additional data necessary to decode them properly. At
>>>> this point the notion of fully decomposing an ID becomes more
>>>> impractical than it's worth, so unabstracting the XY mesh coordinates
>>>> (where 2/3 users were just debug anyway) ends up leaving things a bit
>>>> simpler overall.
>>>>
>>>> Fixes: 60d1504070c2 ("perf/arm-cmn: Support new IP features")
>>>
>>> Does this fix an observable functional issue? It's difficult to tell
>>> what the impact of this is from the commit message.
>>>
>>> i.e. what is the impact:
>>>
>>> * On the CMN-700 programmers model?
>>>
>>> * As observed by a user of the PMU driver?
>>
>> Yes, there are two areas where we functionally depend on decoding the bottom
>> 3 bits of a node ID to the individual port/device numbers. One is just for
>> the debugfs nicety, but the other is in arm_cmn_event_add() for setting
>> input_sel to the appropriate event source.
>
> Cool; and does treating this wrong just result in getting bad values out
> in sysfs/perf, or is it possible that we have something like memory
> corruption?
Yup, it just means we'll configure the DTM to count an event input which
most likely doesn't exist, and thus get 0, plus debugfs looks a little
wonky, e.g.:
1| |
-----+--------+--------
0 | XP #0 | XP #1
| DTC 0 | DTC 0
|........|........
0 | RN-I | RN-D
0| #0 | #0
1| |
1 | HN-T |
0| |
1| |
2 | |
0| #1 |
1| |
3 | |
0| |
1| |
-----+--------+--------
(the HN-T node type still shows up in the right place since that comes
directly from the per-port connect_info registers, but its corresponding
logical ID is looked up via the node ID, thus mistakenly decoded as
belonging to port 2 and output on the wrong line)
>> This was revealed once some hardware turned up with a mix of 3-port and
>> 2-port XPs, and events from a node at port 1 device 0 on a 2-porter were not
>> counting. It took a while to get to the "hang on, why does that ID end in
>> 0x4 not 0x2?" moment...
>
> I see; so we were mislead by the documentation, then saw HW which
> demonstrated this.
>
> It'd be nice to say something like:
>
> The CMN-700 TRM implies that the "extra device ports" configuration is
> global to a CMN instance and the CMN PMU driver currently assumes
> this. Unfortunately this is not correct as this configuration is
> per-XP, and hardware exists where some XPs have extra device ports
> while others do not.
>
> The presence of extra decice ports affects how the bottom 3 bits of a
> node ID map to individual port/device numbers, and when the driver
> misinterprets this, it may expose incorrect information in syfs, and
> arm_cmn_event_add() may configure input_sel incorrectly, resulting in
> incorrect performance counter numbers.
>
> Fix this by reworking node IDs to carry the additional data necessary
> to decode them properly. At this point the notion of fully decomposing
> an ID becomes more impractical than it's worth, so unabstracting the
> XY mesh coordinates (where 2/3 users were just debug anyway) ends up
> leaving things a bit simpler overall.
>
> With that wording (or something very similar), this looks good to me,
> and with that:
Sure, I'll spell out the implications properly - I do tend to forget
that the people reading these won't just be those working on the driver
itself, but also those assessing for backports in general.
Cheers,
Robin.
>
> Acked-by: Mark Rutland <mark.rutland@....com>
>
> Mark.
>
>>> ... and is there anything in the manual that spells out that this is a
>>> per-XP property? I'm struggling to find that in the CMN-700 TRM, as it
>>> seems to talk about "mesh configuration(s) with extra device ports".
>>
>> That's the thing, the only suggestion is where example 3-4 strongly implies
>> it's global by going out of its way to demonstrate the 3-port format on a
>> 2-port XP, despite that being the opposite of reality. (And yes, this has
>> been raised with the documentation folks as well.)
>>
>> Thansk,
>> Robin.
>>
>>>
>>> Mark.
>>>
>>>> Signed-off-by: Robin Murphy <robin.murphy@....com>
>>>> ---
>>>> drivers/perf/arm-cmn.c | 94 ++++++++++++++++++------------------------
>>>> 1 file changed, 40 insertions(+), 54 deletions(-)
>>>>
>>>> diff --git a/drivers/perf/arm-cmn.c b/drivers/perf/arm-cmn.c
>>>> index c932d9d355cf..fd2122a37f22 100644
>>>> --- a/drivers/perf/arm-cmn.c
>>>> +++ b/drivers/perf/arm-cmn.c
>>>> @@ -24,14 +24,6 @@
>>>> #define CMN_NI_NODE_ID GENMASK_ULL(31, 16)
>>>> #define CMN_NI_LOGICAL_ID GENMASK_ULL(47, 32)
>>>> -#define CMN_NODEID_DEVID(reg) ((reg) & 3)
>>>> -#define CMN_NODEID_EXT_DEVID(reg) ((reg) & 1)
>>>> -#define CMN_NODEID_PID(reg) (((reg) >> 2) & 1)
>>>> -#define CMN_NODEID_EXT_PID(reg) (((reg) >> 1) & 3)
>>>> -#define CMN_NODEID_1x1_PID(reg) (((reg) >> 2) & 7)
>>>> -#define CMN_NODEID_X(reg, bits) ((reg) >> (3 + (bits)))
>>>> -#define CMN_NODEID_Y(reg, bits) (((reg) >> 3) & ((1U << (bits)) - 1))
>>>> -
>>>> #define CMN_CHILD_INFO 0x0080
>>>> #define CMN_CI_CHILD_COUNT GENMASK_ULL(15, 0)
>>>> #define CMN_CI_CHILD_PTR_OFFSET GENMASK_ULL(31, 16)
>>>> @@ -280,8 +272,11 @@ struct arm_cmn_node {
>>>> u16 id, logid;
>>>> enum cmn_node_type type;
>>>> + /* XP properties really, but replicated to children for convenience */
>>>> u8 dtm;
>>>> s8 dtc;
>>>> + u8 portid_bits:4;
>>>> + u8 deviceid_bits:4;
>>>> /* DN/HN-F/CXHA */
>>>> struct {
>>>> u8 val : 4;
>>>> @@ -357,49 +352,33 @@ struct arm_cmn {
>>>> static int arm_cmn_hp_state;
>>>> struct arm_cmn_nodeid {
>>>> - u8 x;
>>>> - u8 y;
>>>> u8 port;
>>>> u8 dev;
>>>> };
>>>> static int arm_cmn_xyidbits(const struct arm_cmn *cmn)
>>>> {
>>>> - return fls((cmn->mesh_x - 1) | (cmn->mesh_y - 1) | 2);
>>>> + return fls((cmn->mesh_x - 1) | (cmn->mesh_y - 1));
>>>> }
>>>> -static struct arm_cmn_nodeid arm_cmn_nid(const struct arm_cmn *cmn, u16 id)
>>>> +static struct arm_cmn_nodeid arm_cmn_nid(const struct arm_cmn_node *dn)
>>>> {
>>>> struct arm_cmn_nodeid nid;
>>>> - if (cmn->num_xps == 1) {
>>>> - nid.x = 0;
>>>> - nid.y = 0;
>>>> - nid.port = CMN_NODEID_1x1_PID(id);
>>>> - nid.dev = CMN_NODEID_DEVID(id);
>>>> - } else {
>>>> - int bits = arm_cmn_xyidbits(cmn);
>>>> -
>>>> - nid.x = CMN_NODEID_X(id, bits);
>>>> - nid.y = CMN_NODEID_Y(id, bits);
>>>> - if (cmn->ports_used & 0xc) {
>>>> - nid.port = CMN_NODEID_EXT_PID(id);
>>>> - nid.dev = CMN_NODEID_EXT_DEVID(id);
>>>> - } else {
>>>> - nid.port = CMN_NODEID_PID(id);
>>>> - nid.dev = CMN_NODEID_DEVID(id);
>>>> - }
>>>> - }
>>>> + nid.dev = dn->id & ((1U << dn->deviceid_bits) - 1);
>>>> + nid.port = (dn->id >> dn->deviceid_bits) & ((1U << dn->portid_bits) - 1);
>>>> return nid;
>>>> }
>>>> static struct arm_cmn_node *arm_cmn_node_to_xp(const struct arm_cmn *cmn,
>>>> const struct arm_cmn_node *dn)
>>>> {
>>>> - struct arm_cmn_nodeid nid = arm_cmn_nid(cmn, dn->id);
>>>> - int xp_idx = cmn->mesh_x * nid.y + nid.x;
>>>> + int id = dn->id >> (dn->portid_bits + dn->deviceid_bits);
>>>> + int bits = arm_cmn_xyidbits(cmn);
>>>> + int x = id > bits;
>>>> + int y = id & ((1U << bits) - 1);
>>>> - return cmn->xps + xp_idx;
>>>> + return cmn->xps + cmn->mesh_x * y + x;
>>>> }
>>>> static struct arm_cmn_node *arm_cmn_node(const struct arm_cmn *cmn,
>>>> enum cmn_node_type type)
>>>> @@ -485,13 +464,13 @@ static const char *arm_cmn_device_type(u8 type)
>>>> }
>>>> }
>>>> -static void arm_cmn_show_logid(struct seq_file *s, int x, int y, int p, int d)
>>>> +static void arm_cmn_show_logid(struct seq_file *s, const struct arm_cmn_node *xp, int p, int d)
>>>> {
>>>> struct arm_cmn *cmn = s->private;
>>>> struct arm_cmn_node *dn;
>>>> + u16 id = xp->id | d | (p << xp->deviceid_bits);
>>>> for (dn = cmn->dns; dn->type; dn++) {
>>>> - struct arm_cmn_nodeid nid = arm_cmn_nid(cmn, dn->id);
>>>> int pad = dn->logid < 10;
>>>> if (dn->type == CMN_TYPE_XP)
>>>> @@ -500,7 +479,7 @@ static void arm_cmn_show_logid(struct seq_file *s, int x, int y, int p, int d)
>>>> if (dn->type < CMN_TYPE_HNI)
>>>> continue;
>>>> - if (nid.x != x || nid.y != y || nid.port != p || nid.dev != d)
>>>> + if (dn->id != id)
>>>> continue;
>>>> seq_printf(s, " %*c#%-*d |", pad + 1, ' ', 3 - pad, dn->logid);
>>>> @@ -521,6 +500,7 @@ static int arm_cmn_map_show(struct seq_file *s, void *data)
>>>> y = cmn->mesh_y;
>>>> while (y--) {
>>>> int xp_base = cmn->mesh_x * y;
>>>> + struct arm_cmn_node *xp = cmn->xps + xp_base;
>>>> u8 port[CMN_MAX_PORTS][CMN_MAX_DIMENSION];
>>>> for (x = 0; x < cmn->mesh_x; x++)
>>>> @@ -528,16 +508,14 @@ static int arm_cmn_map_show(struct seq_file *s, void *data)
>>>> seq_printf(s, "\n%-2d |", y);
>>>> for (x = 0; x < cmn->mesh_x; x++) {
>>>> - struct arm_cmn_node *xp = cmn->xps + xp_base + x;
>>>> -
>>>> for (p = 0; p < CMN_MAX_PORTS; p++)
>>>> - port[p][x] = arm_cmn_device_connect_info(cmn, xp, p);
>>>> + port[p][x] = arm_cmn_device_connect_info(cmn, xp + x, p);
>>>> seq_printf(s, " XP #%-3d|", xp_base + x);
>>>> }
>>>> seq_puts(s, "\n |");
>>>> for (x = 0; x < cmn->mesh_x; x++) {
>>>> - s8 dtc = cmn->xps[xp_base + x].dtc;
>>>> + s8 dtc = xp[x].dtc;
>>>> if (dtc < 0)
>>>> seq_puts(s, " DTC ?? |");
>>>> @@ -554,10 +532,10 @@ static int arm_cmn_map_show(struct seq_file *s, void *data)
>>>> seq_puts(s, arm_cmn_device_type(port[p][x]));
>>>> seq_puts(s, "\n 0|");
>>>> for (x = 0; x < cmn->mesh_x; x++)
>>>> - arm_cmn_show_logid(s, x, y, p, 0);
>>>> + arm_cmn_show_logid(s, xp + x, p, 0);
>>>> seq_puts(s, "\n 1|");
>>>> for (x = 0; x < cmn->mesh_x; x++)
>>>> - arm_cmn_show_logid(s, x, y, p, 1);
>>>> + arm_cmn_show_logid(s, xp + x, p, 1);
>>>> }
>>>> seq_puts(s, "\n-----+");
>>>> }
>>>> @@ -1815,10 +1793,7 @@ static int arm_cmn_event_init(struct perf_event *event)
>>>> }
>>>> if (!hw->num_dns) {
>>>> - struct arm_cmn_nodeid nid = arm_cmn_nid(cmn, nodeid);
>>>> -
>>>> - dev_dbg(cmn->dev, "invalid node 0x%x (%d,%d,%d,%d) type 0x%x\n",
>>>> - nodeid, nid.x, nid.y, nid.port, nid.dev, type);
>>>> + dev_dbg(cmn->dev, "invalid node 0x%x type 0x%x\n", nodeid, type);
>>>> return -EINVAL;
>>>> }
>>>> @@ -1921,7 +1896,7 @@ static int arm_cmn_event_add(struct perf_event *event, int flags)
>>>> arm_cmn_claim_wp_idx(dtm, event, d, wp_idx, i);
>>>> writel_relaxed(cfg, dtm->base + CMN_DTM_WPn_CONFIG(wp_idx));
>>>> } else {
>>>> - struct arm_cmn_nodeid nid = arm_cmn_nid(cmn, dn->id);
>>>> + struct arm_cmn_nodeid nid = arm_cmn_nid(dn);
>>>> if (cmn->multi_dtm)
>>>> nid.port %= 2;
>>>> @@ -2168,10 +2143,12 @@ static int arm_cmn_init_dtcs(struct arm_cmn *cmn)
>>>> continue;
>>>> xp = arm_cmn_node_to_xp(cmn, dn);
>>>> + dn->portid_bits = xp->portid_bits;
>>>> + dn->deviceid_bits = xp->deviceid_bits;
>>>> dn->dtc = xp->dtc;
>>>> dn->dtm = xp->dtm;
>>>> if (cmn->multi_dtm)
>>>> - dn->dtm += arm_cmn_nid(cmn, dn->id).port / 2;
>>>> + dn->dtm += arm_cmn_nid(dn).port / 2;
>>>> if (dn->type == CMN_TYPE_DTC) {
>>>> int err = arm_cmn_init_dtc(cmn, dn, dtc_idx++);
>>>> @@ -2341,18 +2318,27 @@ static int arm_cmn_discover(struct arm_cmn *cmn, unsigned int rgn_offset)
>>>> arm_cmn_init_dtm(dtm++, xp, 0);
>>>> /*
>>>> * Keeping track of connected ports will let us filter out
>>>> - * unnecessary XP events easily. We can also reliably infer the
>>>> - * "extra device ports" configuration for the node ID format
>>>> - * from this, since in that case we will see at least one XP
>>>> - * with port 2 connected, for the HN-D.
>>>> + * unnecessary XP events easily, and also infer the per-XP
>>>> + * part of the node ID format.
>>>> */
>>>> for (int p = 0; p < CMN_MAX_PORTS; p++)
>>>> if (arm_cmn_device_connect_info(cmn, xp, p))
>>>> xp_ports |= BIT(p);
>>>> - if (cmn->multi_dtm && (xp_ports & 0xc))
>>>> + if (cmn->num_xps == 1) {
>>>> + xp->portid_bits = 3;
>>>> + xp->deviceid_bits = 2;
>>>> + } else if (xp_ports > 0x3) {
>>>> + xp->portid_bits = 2;
>>>> + xp->deviceid_bits = 1;
>>>> + } else {
>>>> + xp->portid_bits = 1;
>>>> + xp->deviceid_bits = 2;
>>>> + }
>>>> +
>>>> + if (cmn->multi_dtm && (xp_ports > 0x3))
>>>> arm_cmn_init_dtm(dtm++, xp, 1);
>>>> - if (cmn->multi_dtm && (xp_ports & 0x30))
>>>> + if (cmn->multi_dtm && (xp_ports > 0xf))
>>>> arm_cmn_init_dtm(dtm++, xp, 2);
>>>> cmn->ports_used |= xp_ports;
>>>> --
>>>> 2.39.2.101.g768bb238c484.dirty
>>>>
Powered by blists - more mailing lists