* [PATCH v2 1/9] pnv/xive2: XIVE2 Cache Watch, Cache Flush and Sync Injection support
2024-07-15 18:33 [PATCH v2 0/9] XIVE changes for Cache Watch, VSTs, STT and info pic Michael Kowal
@ 2024-07-15 18:33 ` Michael Kowal
2024-07-16 7:28 ` Cédric Le Goater
2024-07-15 18:33 ` [PATCH v2 2/9] pnv/xive2: Structure/define alignment changes Michael Kowal
` (7 subsequent siblings)
8 siblings, 1 reply; 23+ messages in thread
From: Michael Kowal @ 2024-07-15 18:33 UTC (permalink / raw)
To: qemu-devel; +Cc: qemu-ppc, clg, fbarrat, npiggin, milesg
From: Frederic Barrat <fbarrat@linux.ibm.com>
XIVE offers a 'cache watch facility', which allows software to read/update
a potentially cached table entry with no software lock. There's one such
facility in the Virtualization Controller (VC) to update the ESB and END
entries and one in the Presentation Controller (PC) to update the
NVP/NVG/NVC entries.
Each facility has 4 cache watch engines to control the updates and
firmware can request an available engine by querying the hardware
'watch_assign' register of the VC or PC. The engine is then reserved and
is released after the data is updated by reading the 'watch_spec' register
(which also allows to check for a conflict during the update).
If no engine is available, the special value 0xFF is returned and
firmware is expected to repeat the request until an engine becomes
available.
Signed-off-by: Frederic Barrat <fbarrat@linux.ibm.com>
Signed-off-by: Michael Kowal <kowal@linux.vnet.ibm.com>
---
hw/intc/pnv_xive2_regs.h | 90 ++++++++++++++
hw/intc/pnv_xive2.c | 253 +++++++++++++++++++++++++++++++++------
2 files changed, 307 insertions(+), 36 deletions(-)
diff --git a/hw/intc/pnv_xive2_regs.h b/hw/intc/pnv_xive2_regs.h
index 7165dc8704..f8e4a677c6 100644
--- a/hw/intc/pnv_xive2_regs.h
+++ b/hw/intc/pnv_xive2_regs.h
@@ -283,6 +283,15 @@
#define VC_ENDC_SYNC_QUEUE_HARD PPC_BIT(6)
#define VC_QUEUE_COUNT 7
+/* ENDC cache watch assign */
+#define X_VC_ENDC_WATCH_ASSIGN 0x186
+#define VC_ENDC_WATCH_ASSIGN 0x430
+
+/* ENDC configuration register */
+#define X_VC_ENDC_CFG 0x188
+#define VC_ENDC_CFG 0x440
+#define VC_ENDC_CFG_CACHE_WATCH_ASSIGN PPC_BITMASK(32, 35)
+
/* ENDC cache watch specification 0 */
#define X_VC_ENDC_WATCH0_SPEC 0x1A0
#define VC_ENDC_WATCH0_SPEC 0x500
@@ -302,6 +311,42 @@
#define VC_ENDC_WATCH0_DATA2 0x530
#define VC_ENDC_WATCH0_DATA3 0x538
+/* ENDC cache watch 1 */
+#define X_VC_ENDC_WATCH1_SPEC 0x1A8
+#define VC_ENDC_WATCH1_SPEC 0x540
+#define X_VC_ENDC_WATCH1_DATA0 0x1AC
+#define X_VC_ENDC_WATCH1_DATA1 0x1AD
+#define X_VC_ENDC_WATCH1_DATA2 0x1AE
+#define X_VC_ENDC_WATCH1_DATA3 0x1AF
+#define VC_ENDC_WATCH1_DATA0 0x560
+#define VC_ENDC_WATCH1_DATA1 0x568
+#define VC_ENDC_WATCH1_DATA2 0x570
+#define VC_ENDC_WATCH1_DATA3 0x578
+
+/* ENDC cache watch 2 */
+#define X_VC_ENDC_WATCH2_SPEC 0x1B0
+#define VC_ENDC_WATCH2_SPEC 0x580
+#define X_VC_ENDC_WATCH2_DATA0 0x1B4
+#define X_VC_ENDC_WATCH2_DATA1 0x1B5
+#define X_VC_ENDC_WATCH2_DATA2 0x1B6
+#define X_VC_ENDC_WATCH2_DATA3 0x1B7
+#define VC_ENDC_WATCH2_DATA0 0x5A0
+#define VC_ENDC_WATCH2_DATA1 0x5A8
+#define VC_ENDC_WATCH2_DATA2 0x5B0
+#define VC_ENDC_WATCH2_DATA3 0x5B8
+
+/* ENDC cache watch 3 */
+#define X_VC_ENDC_WATCH3_SPEC 0x1B8
+#define VC_ENDC_WATCH3_SPEC 0x5C0
+#define X_VC_ENDC_WATCH3_DATA0 0x1BC
+#define X_VC_ENDC_WATCH3_DATA1 0x1BD
+#define X_VC_ENDC_WATCH3_DATA2 0x1BE
+#define X_VC_ENDC_WATCH3_DATA3 0x1BF
+#define VC_ENDC_WATCH3_DATA0 0x5E0
+#define VC_ENDC_WATCH3_DATA1 0x5E8
+#define VC_ENDC_WATCH3_DATA2 0x5F0
+#define VC_ENDC_WATCH3_DATA3 0x5F8
+
/*
* PC LSB1
*/
@@ -358,6 +403,15 @@
#define PC_NXC_FLUSH_POLL_BLOCK_ID_MASK PPC_BITMASK(36, 39)
#define PC_NXC_FLUSH_POLL_OFFSET_MASK PPC_BITMASK(40, 63) /* 24-bit */
+/* NxC Cache watch assign */
+#define X_PC_NXC_WATCH_ASSIGN 0x286
+#define PC_NXC_WATCH_ASSIGN 0x430
+
+/* NxC Proc config */
+#define X_PC_NXC_PROC_CONFIG 0x28A
+#define PC_NXC_PROC_CONFIG 0x450
+#define PC_NXC_PROC_CONFIG_WATCH_ASSIGN PPC_BITMASK(0, 3)
+
/* NxC Cache Watch 0 Specification */
#define X_PC_NXC_WATCH0_SPEC 0x2A0
#define PC_NXC_WATCH0_SPEC 0x500
@@ -381,6 +435,42 @@
#define PC_NXC_WATCH0_DATA2 0x530
#define PC_NXC_WATCH0_DATA3 0x538
+/* NxC Cache Watch 1 */
+#define X_PC_NXC_WATCH1_SPEC 0x2A8
+#define PC_NXC_WATCH1_SPEC 0x540
+#define X_PC_NXC_WATCH1_DATA0 0x2AC
+#define X_PC_NXC_WATCH1_DATA1 0x2AD
+#define X_PC_NXC_WATCH1_DATA2 0x2AE
+#define X_PC_NXC_WATCH1_DATA3 0x2AF
+#define PC_NXC_WATCH1_DATA0 0x560
+#define PC_NXC_WATCH1_DATA1 0x568
+#define PC_NXC_WATCH1_DATA2 0x570
+#define PC_NXC_WATCH1_DATA3 0x578
+
+/* NxC Cache Watch 2 */
+#define X_PC_NXC_WATCH2_SPEC 0x2B0
+#define PC_NXC_WATCH2_SPEC 0x580
+#define X_PC_NXC_WATCH2_DATA0 0x2B4
+#define X_PC_NXC_WATCH2_DATA1 0x2B5
+#define X_PC_NXC_WATCH2_DATA2 0x2B6
+#define X_PC_NXC_WATCH2_DATA3 0x2B7
+#define PC_NXC_WATCH2_DATA0 0x5A0
+#define PC_NXC_WATCH2_DATA1 0x5A8
+#define PC_NXC_WATCH2_DATA2 0x5B0
+#define PC_NXC_WATCH2_DATA3 0x5B8
+
+/* NxC Cache Watch 3 */
+#define X_PC_NXC_WATCH3_SPEC 0x2B8
+#define PC_NXC_WATCH3_SPEC 0x5C0
+#define X_PC_NXC_WATCH3_DATA0 0x2BC
+#define X_PC_NXC_WATCH3_DATA1 0x2BD
+#define X_PC_NXC_WATCH3_DATA2 0x2BE
+#define X_PC_NXC_WATCH3_DATA3 0x2BF
+#define PC_NXC_WATCH3_DATA0 0x5E0
+#define PC_NXC_WATCH3_DATA1 0x5E8
+#define PC_NXC_WATCH3_DATA2 0x5F0
+#define PC_NXC_WATCH3_DATA3 0x5F8
+
/*
* TCTXT Registers
*/
diff --git a/hw/intc/pnv_xive2.c b/hw/intc/pnv_xive2.c
index 2fb4fa29d4..f6a735cca5 100644
--- a/hw/intc/pnv_xive2.c
+++ b/hw/intc/pnv_xive2.c
@@ -329,40 +329,48 @@ static int pnv_xive2_write_end(Xive2Router *xrtr, uint8_t blk, uint32_t idx,
word_number);
}
-static int pnv_xive2_end_update(PnvXive2 *xive)
+static int pnv_xive2_end_update(PnvXive2 *xive, uint8_t watch_engine)
{
- uint8_t blk = GETFIELD(VC_ENDC_WATCH_BLOCK_ID,
- xive->vc_regs[(VC_ENDC_WATCH0_SPEC >> 3)]);
- uint32_t idx = GETFIELD(VC_ENDC_WATCH_INDEX,
- xive->vc_regs[(VC_ENDC_WATCH0_SPEC >> 3)]);
- int i;
+ uint8_t blk;
+ uint32_t idx;
+ int i, spec_reg, data_reg;
uint64_t endc_watch[4];
+ assert(watch_engine < ARRAY_SIZE(endc_watch));
+
+ spec_reg = (VC_ENDC_WATCH0_SPEC + watch_engine * 0x40) >> 3;
+ data_reg = (VC_ENDC_WATCH0_DATA0 + watch_engine * 0x40) >> 3;
+ blk = GETFIELD(VC_ENDC_WATCH_BLOCK_ID, xive->vc_regs[spec_reg]);
+ idx = GETFIELD(VC_ENDC_WATCH_INDEX, xive->vc_regs[spec_reg]);
+
for (i = 0; i < ARRAY_SIZE(endc_watch); i++) {
- endc_watch[i] =
- cpu_to_be64(xive->vc_regs[(VC_ENDC_WATCH0_DATA0 >> 3) + i]);
+ endc_watch[i] = cpu_to_be64(xive->vc_regs[data_reg + i]);
}
return pnv_xive2_vst_write(xive, VST_END, blk, idx, endc_watch,
XIVE_VST_WORD_ALL);
}
-static void pnv_xive2_end_cache_load(PnvXive2 *xive)
+static void pnv_xive2_end_cache_load(PnvXive2 *xive, uint8_t watch_engine)
{
- uint8_t blk = GETFIELD(VC_ENDC_WATCH_BLOCK_ID,
- xive->vc_regs[(VC_ENDC_WATCH0_SPEC >> 3)]);
- uint32_t idx = GETFIELD(VC_ENDC_WATCH_INDEX,
- xive->vc_regs[(VC_ENDC_WATCH0_SPEC >> 3)]);
+ uint8_t blk;
+ uint32_t idx;
uint64_t endc_watch[4] = { 0 };
- int i;
+ int i, spec_reg, data_reg;
+
+ assert(watch_engine < ARRAY_SIZE(endc_watch));
+
+ spec_reg = (VC_ENDC_WATCH0_SPEC + watch_engine * 0x40) >> 3;
+ data_reg = (VC_ENDC_WATCH0_DATA0 + watch_engine * 0x40) >> 3;
+ blk = GETFIELD(VC_ENDC_WATCH_BLOCK_ID, xive->vc_regs[spec_reg]);
+ idx = GETFIELD(VC_ENDC_WATCH_INDEX, xive->vc_regs[spec_reg]);
if (pnv_xive2_vst_read(xive, VST_END, blk, idx, endc_watch)) {
xive2_error(xive, "VST: no END entry %x/%x !?", blk, idx);
}
for (i = 0; i < ARRAY_SIZE(endc_watch); i++) {
- xive->vc_regs[(VC_ENDC_WATCH0_DATA0 >> 3) + i] =
- be64_to_cpu(endc_watch[i]);
+ xive->vc_regs[data_reg + i] = be64_to_cpu(endc_watch[i]);
}
}
@@ -379,40 +387,50 @@ static int pnv_xive2_write_nvp(Xive2Router *xrtr, uint8_t blk, uint32_t idx,
word_number);
}
-static int pnv_xive2_nvp_update(PnvXive2 *xive)
+static int pnv_xive2_nvp_update(PnvXive2 *xive, uint8_t watch_engine)
{
- uint8_t blk = GETFIELD(PC_NXC_WATCH_BLOCK_ID,
- xive->pc_regs[(PC_NXC_WATCH0_SPEC >> 3)]);
- uint32_t idx = GETFIELD(PC_NXC_WATCH_INDEX,
- xive->pc_regs[(PC_NXC_WATCH0_SPEC >> 3)]);
- int i;
+ uint8_t blk;
+ uint32_t idx;
+ int i, spec_reg, data_reg;
uint64_t nxc_watch[4];
+ if (watch_engine > 3) {
+ return -1;
+ }
+ spec_reg = (PC_NXC_WATCH0_SPEC + watch_engine * 0x40) >> 3;
+ data_reg = (PC_NXC_WATCH0_DATA0 + watch_engine * 0x40) >> 3;
+ blk = GETFIELD(PC_NXC_WATCH_BLOCK_ID, xive->pc_regs[spec_reg]);
+ idx = GETFIELD(PC_NXC_WATCH_INDEX, xive->pc_regs[spec_reg]);
+
for (i = 0; i < ARRAY_SIZE(nxc_watch); i++) {
- nxc_watch[i] =
- cpu_to_be64(xive->pc_regs[(PC_NXC_WATCH0_DATA0 >> 3) + i]);
+ nxc_watch[i] = cpu_to_be64(xive->pc_regs[data_reg + i]);
}
return pnv_xive2_vst_write(xive, VST_NVP, blk, idx, nxc_watch,
XIVE_VST_WORD_ALL);
}
-static void pnv_xive2_nvp_cache_load(PnvXive2 *xive)
+static void pnv_xive2_nvp_cache_load(PnvXive2 *xive, uint8_t watch_engine)
{
- uint8_t blk = GETFIELD(PC_NXC_WATCH_BLOCK_ID,
- xive->pc_regs[(PC_NXC_WATCH0_SPEC >> 3)]);
- uint32_t idx = GETFIELD(PC_NXC_WATCH_INDEX,
- xive->pc_regs[(PC_NXC_WATCH0_SPEC >> 3)]);
+ uint8_t blk;
+ uint32_t idx;
uint64_t nxc_watch[4] = { 0 };
- int i;
+ int i, spec_reg, data_reg;
+
+ if (watch_engine > 3) {
+ return;
+ }
+ spec_reg = (PC_NXC_WATCH0_SPEC + watch_engine * 0x40) >> 3;
+ data_reg = (PC_NXC_WATCH0_DATA0 + watch_engine * 0x40) >> 3;
+ blk = GETFIELD(PC_NXC_WATCH_BLOCK_ID, xive->pc_regs[spec_reg]);
+ idx = GETFIELD(PC_NXC_WATCH_INDEX, xive->pc_regs[spec_reg]);
if (pnv_xive2_vst_read(xive, VST_NVP, blk, idx, nxc_watch)) {
xive2_error(xive, "VST: no NVP entry %x/%x !?", blk, idx);
}
for (i = 0; i < ARRAY_SIZE(nxc_watch); i++) {
- xive->pc_regs[(PC_NXC_WATCH0_DATA0 >> 3) + i] =
- be64_to_cpu(nxc_watch[i]);
+ xive->pc_regs[data_reg + i] = be64_to_cpu(nxc_watch[i]);
}
}
@@ -964,12 +982,70 @@ static const MemoryRegionOps pnv_xive2_ic_cq_ops = {
},
};
+static uint8_t pnv_xive2_cache_watch_assign(uint64_t engine_mask,
+ uint64_t *state)
+{
+ uint8_t val = 0xFF;
+ int i;
+
+ for (i = 3; i >= 0; i--) {
+ if (BIT(i) & engine_mask) {
+ if (!(BIT(i) & *state)) {
+ *state |= BIT(i);
+ val = 3 - i;
+ break;
+ }
+ }
+ }
+ return val;
+}
+
+static void pnv_xive2_cache_watch_release(uint64_t *state, uint8_t watch_engine)
+{
+ uint8_t engine_bit = 3 - watch_engine;
+
+ if (*state & BIT(engine_bit)) {
+ *state &= ~BIT(engine_bit);
+ }
+}
+
+static uint8_t pnv_xive2_endc_cache_watch_assign(PnvXive2 *xive)
+{
+ uint64_t engine_mask = GETFIELD(VC_ENDC_CFG_CACHE_WATCH_ASSIGN,
+ xive->vc_regs[VC_ENDC_CFG >> 3]);
+ uint64_t state = xive->vc_regs[VC_ENDC_WATCH_ASSIGN >> 3];
+ uint8_t val;
+
+ /*
+ * We keep track of which engines are currently busy in the
+ * VC_ENDC_WATCH_ASSIGN register directly. When the firmware reads
+ * the register, we don't return its value but the ID of an engine
+ * it can use.
+ * There are 4 engines. 0xFF means no engine is available.
+ */
+ val = pnv_xive2_cache_watch_assign(engine_mask, &state);
+ if (val != 0xFF) {
+ xive->vc_regs[VC_ENDC_WATCH_ASSIGN >> 3] = state;
+ }
+ return val;
+}
+
+static void pnv_xive2_endc_cache_watch_release(PnvXive2 *xive,
+ uint8_t watch_engine)
+{
+ uint64_t state = xive->vc_regs[VC_ENDC_WATCH_ASSIGN >> 3];
+
+ pnv_xive2_cache_watch_release(&state, watch_engine);
+ xive->vc_regs[VC_ENDC_WATCH_ASSIGN >> 3] = state;
+}
+
static uint64_t pnv_xive2_ic_vc_read(void *opaque, hwaddr offset,
unsigned size)
{
PnvXive2 *xive = PNV_XIVE2(opaque);
uint64_t val = 0;
uint32_t reg = offset >> 3;
+ uint8_t watch_engine;
switch (offset) {
/*
@@ -1000,24 +1076,44 @@ static uint64_t pnv_xive2_ic_vc_read(void *opaque, hwaddr offset,
val = xive->vc_regs[reg];
break;
+ case VC_ENDC_WATCH_ASSIGN:
+ val = pnv_xive2_endc_cache_watch_assign(xive);
+ break;
+
+ case VC_ENDC_CFG:
+ val = xive->vc_regs[reg];
+ break;
+
/*
* END cache updates
*/
case VC_ENDC_WATCH0_SPEC:
+ case VC_ENDC_WATCH1_SPEC:
+ case VC_ENDC_WATCH2_SPEC:
+ case VC_ENDC_WATCH3_SPEC:
+ watch_engine = (offset - VC_ENDC_WATCH0_SPEC) >> 6;
xive->vc_regs[reg] &= ~(VC_ENDC_WATCH_FULL | VC_ENDC_WATCH_CONFLICT);
+ pnv_xive2_endc_cache_watch_release(xive, watch_engine);
val = xive->vc_regs[reg];
break;
case VC_ENDC_WATCH0_DATA0:
+ case VC_ENDC_WATCH1_DATA0:
+ case VC_ENDC_WATCH2_DATA0:
+ case VC_ENDC_WATCH3_DATA0:
/*
* Load DATA registers from cache with data requested by the
* SPEC register
*/
- pnv_xive2_end_cache_load(xive);
+ watch_engine = (offset - VC_ENDC_WATCH0_DATA0) >> 6;
+ pnv_xive2_end_cache_load(xive, watch_engine);
val = xive->vc_regs[reg];
break;
case VC_ENDC_WATCH0_DATA1 ... VC_ENDC_WATCH0_DATA3:
+ case VC_ENDC_WATCH1_DATA1 ... VC_ENDC_WATCH1_DATA3:
+ case VC_ENDC_WATCH2_DATA1 ... VC_ENDC_WATCH2_DATA3:
+ case VC_ENDC_WATCH3_DATA1 ... VC_ENDC_WATCH3_DATA3:
val = xive->vc_regs[reg];
break;
@@ -1063,6 +1159,7 @@ static void pnv_xive2_ic_vc_write(void *opaque, hwaddr offset,
{
PnvXive2 *xive = PNV_XIVE2(opaque);
uint32_t reg = offset >> 3;
+ uint8_t watch_engine;
switch (offset) {
/*
@@ -1095,19 +1192,32 @@ static void pnv_xive2_ic_vc_write(void *opaque, hwaddr offset,
/* EAS update */
break;
+ case VC_ENDC_CFG:
+ break;
+
/*
* END cache updates
*/
case VC_ENDC_WATCH0_SPEC:
+ case VC_ENDC_WATCH1_SPEC:
+ case VC_ENDC_WATCH2_SPEC:
+ case VC_ENDC_WATCH3_SPEC:
val &= ~VC_ENDC_WATCH_CONFLICT; /* HW will set this bit */
break;
case VC_ENDC_WATCH0_DATA1 ... VC_ENDC_WATCH0_DATA3:
+ case VC_ENDC_WATCH1_DATA1 ... VC_ENDC_WATCH1_DATA3:
+ case VC_ENDC_WATCH2_DATA1 ... VC_ENDC_WATCH2_DATA3:
+ case VC_ENDC_WATCH3_DATA1 ... VC_ENDC_WATCH3_DATA3:
break;
case VC_ENDC_WATCH0_DATA0:
+ case VC_ENDC_WATCH1_DATA0:
+ case VC_ENDC_WATCH2_DATA0:
+ case VC_ENDC_WATCH3_DATA0:
/* writing to DATA0 triggers the cache write */
+ watch_engine = (offset - VC_ENDC_WATCH0_DATA0) >> 6;
xive->vc_regs[reg] = val;
- pnv_xive2_end_update(xive);
+ pnv_xive2_end_update(xive, watch_engine);
break;
@@ -1157,12 +1267,43 @@ static const MemoryRegionOps pnv_xive2_ic_vc_ops = {
},
};
+static uint8_t pnv_xive2_nxc_cache_watch_assign(PnvXive2 *xive)
+{
+ uint64_t engine_mask = GETFIELD(PC_NXC_PROC_CONFIG_WATCH_ASSIGN,
+ xive->pc_regs[PC_NXC_PROC_CONFIG >> 3]);
+ uint64_t state = xive->pc_regs[PC_NXC_WATCH_ASSIGN >> 3];
+ uint8_t val;
+
+ /*
+ * We keep track of which engines are currently busy in the
+ * PC_NXC_WATCH_ASSIGN register directly. When the firmware reads
+ * the register, we don't return its value but the ID of an engine
+ * it can use.
+ * There are 4 engines. 0xFF means no engine is available.
+ */
+ val = pnv_xive2_cache_watch_assign(engine_mask, &state);
+ if (val != 0xFF) {
+ xive->pc_regs[PC_NXC_WATCH_ASSIGN >> 3] = state;
+ }
+ return val;
+}
+
+static void pnv_xive2_nxc_cache_watch_release(PnvXive2 *xive,
+ uint8_t watch_engine)
+{
+ uint64_t state = xive->pc_regs[PC_NXC_WATCH_ASSIGN >> 3];
+
+ pnv_xive2_cache_watch_release(&state, watch_engine);
+ xive->pc_regs[PC_NXC_WATCH_ASSIGN >> 3] = state;
+}
+
static uint64_t pnv_xive2_ic_pc_read(void *opaque, hwaddr offset,
unsigned size)
{
PnvXive2 *xive = PNV_XIVE2(opaque);
uint64_t val = -1;
uint32_t reg = offset >> 3;
+ uint8_t watch_engine;
switch (offset) {
/*
@@ -1173,24 +1314,44 @@ static uint64_t pnv_xive2_ic_pc_read(void *opaque, hwaddr offset,
val = xive->pc_regs[reg];
break;
+ case PC_NXC_WATCH_ASSIGN:
+ val = pnv_xive2_nxc_cache_watch_assign(xive);
+ break;
+
+ case PC_NXC_PROC_CONFIG:
+ val = xive->pc_regs[reg];
+ break;
+
/*
* cache updates
*/
case PC_NXC_WATCH0_SPEC:
+ case PC_NXC_WATCH1_SPEC:
+ case PC_NXC_WATCH2_SPEC:
+ case PC_NXC_WATCH3_SPEC:
+ watch_engine = (offset - PC_NXC_WATCH0_SPEC) >> 6;
xive->pc_regs[reg] &= ~(PC_NXC_WATCH_FULL | PC_NXC_WATCH_CONFLICT);
+ pnv_xive2_nxc_cache_watch_release(xive, watch_engine);
val = xive->pc_regs[reg];
break;
case PC_NXC_WATCH0_DATA0:
+ case PC_NXC_WATCH1_DATA0:
+ case PC_NXC_WATCH2_DATA0:
+ case PC_NXC_WATCH3_DATA0:
/*
* Load DATA registers from cache with data requested by the
* SPEC register
*/
- pnv_xive2_nvp_cache_load(xive);
+ watch_engine = (offset - PC_NXC_WATCH0_DATA0) >> 6;
+ pnv_xive2_nvp_cache_load(xive, watch_engine);
val = xive->pc_regs[reg];
break;
case PC_NXC_WATCH0_DATA1 ... PC_NXC_WATCH0_DATA3:
+ case PC_NXC_WATCH1_DATA1 ... PC_NXC_WATCH1_DATA3:
+ case PC_NXC_WATCH2_DATA1 ... PC_NXC_WATCH2_DATA3:
+ case PC_NXC_WATCH3_DATA1 ... PC_NXC_WATCH3_DATA3:
val = xive->pc_regs[reg];
break;
@@ -1219,6 +1380,7 @@ static void pnv_xive2_ic_pc_write(void *opaque, hwaddr offset,
{
PnvXive2 *xive = PNV_XIVE2(opaque);
uint32_t reg = offset >> 3;
+ uint8_t watch_engine;
switch (offset) {
@@ -1231,19 +1393,32 @@ static void pnv_xive2_ic_pc_write(void *opaque, hwaddr offset,
case PC_VSD_TABLE_DATA:
break;
+ case PC_NXC_PROC_CONFIG:
+ break;
+
/*
* cache updates
*/
case PC_NXC_WATCH0_SPEC:
+ case PC_NXC_WATCH1_SPEC:
+ case PC_NXC_WATCH2_SPEC:
+ case PC_NXC_WATCH3_SPEC:
val &= ~PC_NXC_WATCH_CONFLICT; /* HW will set this bit */
break;
case PC_NXC_WATCH0_DATA1 ... PC_NXC_WATCH0_DATA3:
+ case PC_NXC_WATCH1_DATA1 ... PC_NXC_WATCH1_DATA3:
+ case PC_NXC_WATCH2_DATA1 ... PC_NXC_WATCH2_DATA3:
+ case PC_NXC_WATCH3_DATA1 ... PC_NXC_WATCH3_DATA3:
break;
case PC_NXC_WATCH0_DATA0:
+ case PC_NXC_WATCH1_DATA0:
+ case PC_NXC_WATCH2_DATA0:
+ case PC_NXC_WATCH3_DATA0:
/* writing to DATA0 triggers the cache write */
+ watch_engine = (offset - PC_NXC_WATCH0_DATA0) >> 6;
xive->pc_regs[reg] = val;
- pnv_xive2_nvp_update(xive);
+ pnv_xive2_nvp_update(xive, watch_engine);
break;
/* case PC_NXC_FLUSH_CTRL: */
@@ -1814,6 +1989,12 @@ static void pnv_xive2_reset(void *dev)
xive->cq_regs[CQ_XIVE_CFG >> 3] |=
SETFIELD(CQ_XIVE_CFG_HYP_HARD_BLOCK_ID, 0ull, xive->chip->chip_id);
+ /* VC and PC cache watch assign mechanism */
+ xive->vc_regs[VC_ENDC_CFG >> 3] =
+ SETFIELD(VC_ENDC_CFG_CACHE_WATCH_ASSIGN, 0ull, 0b0111);
+ xive->pc_regs[PC_NXC_PROC_CONFIG >> 3] =
+ SETFIELD(PC_NXC_PROC_CONFIG_WATCH_ASSIGN, 0ull, 0b0111);
+
/* Set default page size to 64k */
xive->ic_shift = xive->esb_shift = xive->end_shift = 16;
xive->nvc_shift = xive->nvpg_shift = xive->tm_shift = 16;
--
2.43.0
^ permalink raw reply related [flat|nested] 23+ messages in thread* Re: [PATCH v2 1/9] pnv/xive2: XIVE2 Cache Watch, Cache Flush and Sync Injection support
2024-07-15 18:33 ` [PATCH v2 1/9] pnv/xive2: XIVE2 Cache Watch, Cache Flush and Sync Injection support Michael Kowal
@ 2024-07-16 7:28 ` Cédric Le Goater
2024-07-16 15:32 ` Mike Kowal
0 siblings, 1 reply; 23+ messages in thread
From: Cédric Le Goater @ 2024-07-16 7:28 UTC (permalink / raw)
To: Michael Kowal, qemu-devel; +Cc: qemu-ppc, fbarrat, npiggin, milesg
On 7/15/24 20:33, Michael Kowal wrote:
> From: Frederic Barrat <fbarrat@linux.ibm.com>
>
> XIVE offers a 'cache watch facility', which allows software to read/update
> a potentially cached table entry with no software lock. There's one such
> facility in the Virtualization Controller (VC) to update the ESB and END
> entries and one in the Presentation Controller (PC) to update the
> NVP/NVG/NVC entries.
>
> Each facility has 4 cache watch engines to control the updates and
> firmware can request an available engine by querying the hardware
> 'watch_assign' register of the VC or PC. The engine is then reserved and
> is released after the data is updated by reading the 'watch_spec' register
> (which also allows to check for a conflict during the update).
> If no engine is available, the special value 0xFF is returned and
> firmware is expected to repeat the request until an engine becomes
> available.
>
> Signed-off-by: Frederic Barrat <fbarrat@linux.ibm.com>
> Signed-off-by: Michael Kowal <kowal@linux.vnet.ibm.com>
> ---
> hw/intc/pnv_xive2_regs.h | 90 ++++++++++++++
> hw/intc/pnv_xive2.c | 253 +++++++++++++++++++++++++++++++++------
> 2 files changed, 307 insertions(+), 36 deletions(-)
>
> diff --git a/hw/intc/pnv_xive2_regs.h b/hw/intc/pnv_xive2_regs.h
> index 7165dc8704..f8e4a677c6 100644
> --- a/hw/intc/pnv_xive2_regs.h
> +++ b/hw/intc/pnv_xive2_regs.h
> @@ -283,6 +283,15 @@
> #define VC_ENDC_SYNC_QUEUE_HARD PPC_BIT(6)
> #define VC_QUEUE_COUNT 7
>
> +/* ENDC cache watch assign */
> +#define X_VC_ENDC_WATCH_ASSIGN 0x186
> +#define VC_ENDC_WATCH_ASSIGN 0x430
> +
> +/* ENDC configuration register */
> +#define X_VC_ENDC_CFG 0x188
> +#define VC_ENDC_CFG 0x440
> +#define VC_ENDC_CFG_CACHE_WATCH_ASSIGN PPC_BITMASK(32, 35)
> +
> /* ENDC cache watch specification 0 */
> #define X_VC_ENDC_WATCH0_SPEC 0x1A0
> #define VC_ENDC_WATCH0_SPEC 0x500
> @@ -302,6 +311,42 @@
> #define VC_ENDC_WATCH0_DATA2 0x530
> #define VC_ENDC_WATCH0_DATA3 0x538
>
> +/* ENDC cache watch 1 */
> +#define X_VC_ENDC_WATCH1_SPEC 0x1A8
> +#define VC_ENDC_WATCH1_SPEC 0x540
> +#define X_VC_ENDC_WATCH1_DATA0 0x1AC
> +#define X_VC_ENDC_WATCH1_DATA1 0x1AD
> +#define X_VC_ENDC_WATCH1_DATA2 0x1AE
> +#define X_VC_ENDC_WATCH1_DATA3 0x1AF
> +#define VC_ENDC_WATCH1_DATA0 0x560
> +#define VC_ENDC_WATCH1_DATA1 0x568
> +#define VC_ENDC_WATCH1_DATA2 0x570
> +#define VC_ENDC_WATCH1_DATA3 0x578
> +
> +/* ENDC cache watch 2 */
> +#define X_VC_ENDC_WATCH2_SPEC 0x1B0
> +#define VC_ENDC_WATCH2_SPEC 0x580
> +#define X_VC_ENDC_WATCH2_DATA0 0x1B4
> +#define X_VC_ENDC_WATCH2_DATA1 0x1B5
> +#define X_VC_ENDC_WATCH2_DATA2 0x1B6
> +#define X_VC_ENDC_WATCH2_DATA3 0x1B7
> +#define VC_ENDC_WATCH2_DATA0 0x5A0
> +#define VC_ENDC_WATCH2_DATA1 0x5A8
> +#define VC_ENDC_WATCH2_DATA2 0x5B0
> +#define VC_ENDC_WATCH2_DATA3 0x5B8
> +
> +/* ENDC cache watch 3 */
> +#define X_VC_ENDC_WATCH3_SPEC 0x1B8
> +#define VC_ENDC_WATCH3_SPEC 0x5C0
> +#define X_VC_ENDC_WATCH3_DATA0 0x1BC
> +#define X_VC_ENDC_WATCH3_DATA1 0x1BD
> +#define X_VC_ENDC_WATCH3_DATA2 0x1BE
> +#define X_VC_ENDC_WATCH3_DATA3 0x1BF
> +#define VC_ENDC_WATCH3_DATA0 0x5E0
> +#define VC_ENDC_WATCH3_DATA1 0x5E8
> +#define VC_ENDC_WATCH3_DATA2 0x5F0
> +#define VC_ENDC_WATCH3_DATA3 0x5F8
> +
> /*
> * PC LSB1
> */
> @@ -358,6 +403,15 @@
> #define PC_NXC_FLUSH_POLL_BLOCK_ID_MASK PPC_BITMASK(36, 39)
> #define PC_NXC_FLUSH_POLL_OFFSET_MASK PPC_BITMASK(40, 63) /* 24-bit */
>
> +/* NxC Cache watch assign */
> +#define X_PC_NXC_WATCH_ASSIGN 0x286
> +#define PC_NXC_WATCH_ASSIGN 0x430
> +
> +/* NxC Proc config */
> +#define X_PC_NXC_PROC_CONFIG 0x28A
> +#define PC_NXC_PROC_CONFIG 0x450
> +#define PC_NXC_PROC_CONFIG_WATCH_ASSIGN PPC_BITMASK(0, 3)
> +
> /* NxC Cache Watch 0 Specification */
> #define X_PC_NXC_WATCH0_SPEC 0x2A0
> #define PC_NXC_WATCH0_SPEC 0x500
> @@ -381,6 +435,42 @@
> #define PC_NXC_WATCH0_DATA2 0x530
> #define PC_NXC_WATCH0_DATA3 0x538
>
> +/* NxC Cache Watch 1 */
> +#define X_PC_NXC_WATCH1_SPEC 0x2A8
> +#define PC_NXC_WATCH1_SPEC 0x540
> +#define X_PC_NXC_WATCH1_DATA0 0x2AC
> +#define X_PC_NXC_WATCH1_DATA1 0x2AD
> +#define X_PC_NXC_WATCH1_DATA2 0x2AE
> +#define X_PC_NXC_WATCH1_DATA3 0x2AF
> +#define PC_NXC_WATCH1_DATA0 0x560
> +#define PC_NXC_WATCH1_DATA1 0x568
> +#define PC_NXC_WATCH1_DATA2 0x570
> +#define PC_NXC_WATCH1_DATA3 0x578
> +
> +/* NxC Cache Watch 2 */
> +#define X_PC_NXC_WATCH2_SPEC 0x2B0
> +#define PC_NXC_WATCH2_SPEC 0x580
> +#define X_PC_NXC_WATCH2_DATA0 0x2B4
> +#define X_PC_NXC_WATCH2_DATA1 0x2B5
> +#define X_PC_NXC_WATCH2_DATA2 0x2B6
> +#define X_PC_NXC_WATCH2_DATA3 0x2B7
> +#define PC_NXC_WATCH2_DATA0 0x5A0
> +#define PC_NXC_WATCH2_DATA1 0x5A8
> +#define PC_NXC_WATCH2_DATA2 0x5B0
> +#define PC_NXC_WATCH2_DATA3 0x5B8
> +
> +/* NxC Cache Watch 3 */
> +#define X_PC_NXC_WATCH3_SPEC 0x2B8
> +#define PC_NXC_WATCH3_SPEC 0x5C0
> +#define X_PC_NXC_WATCH3_DATA0 0x2BC
> +#define X_PC_NXC_WATCH3_DATA1 0x2BD
> +#define X_PC_NXC_WATCH3_DATA2 0x2BE
> +#define X_PC_NXC_WATCH3_DATA3 0x2BF
> +#define PC_NXC_WATCH3_DATA0 0x5E0
> +#define PC_NXC_WATCH3_DATA1 0x5E8
> +#define PC_NXC_WATCH3_DATA2 0x5F0
> +#define PC_NXC_WATCH3_DATA3 0x5F8
> +
> /*
> * TCTXT Registers
> */
> diff --git a/hw/intc/pnv_xive2.c b/hw/intc/pnv_xive2.c
> index 2fb4fa29d4..f6a735cca5 100644
> --- a/hw/intc/pnv_xive2.c
> +++ b/hw/intc/pnv_xive2.c
> @@ -329,40 +329,48 @@ static int pnv_xive2_write_end(Xive2Router *xrtr, uint8_t blk, uint32_t idx,
> word_number);
> }
>
> -static int pnv_xive2_end_update(PnvXive2 *xive)
> +static int pnv_xive2_end_update(PnvXive2 *xive, uint8_t watch_engine)
> {
> - uint8_t blk = GETFIELD(VC_ENDC_WATCH_BLOCK_ID,
> - xive->vc_regs[(VC_ENDC_WATCH0_SPEC >> 3)]);
> - uint32_t idx = GETFIELD(VC_ENDC_WATCH_INDEX,
> - xive->vc_regs[(VC_ENDC_WATCH0_SPEC >> 3)]);
> - int i;
> + uint8_t blk;
> + uint32_t idx;
> + int i, spec_reg, data_reg;
> uint64_t endc_watch[4];
>
> + assert(watch_engine < ARRAY_SIZE(endc_watch));
> +
> + spec_reg = (VC_ENDC_WATCH0_SPEC + watch_engine * 0x40) >> 3;
> + data_reg = (VC_ENDC_WATCH0_DATA0 + watch_engine * 0x40) >> 3;
> + blk = GETFIELD(VC_ENDC_WATCH_BLOCK_ID, xive->vc_regs[spec_reg]);
> + idx = GETFIELD(VC_ENDC_WATCH_INDEX, xive->vc_regs[spec_reg]);
> +
> for (i = 0; i < ARRAY_SIZE(endc_watch); i++) {
> - endc_watch[i] =
> - cpu_to_be64(xive->vc_regs[(VC_ENDC_WATCH0_DATA0 >> 3) + i]);
> + endc_watch[i] = cpu_to_be64(xive->vc_regs[data_reg + i]);
> }
>
> return pnv_xive2_vst_write(xive, VST_END, blk, idx, endc_watch,
> XIVE_VST_WORD_ALL);
> }
>
> -static void pnv_xive2_end_cache_load(PnvXive2 *xive)
> +static void pnv_xive2_end_cache_load(PnvXive2 *xive, uint8_t watch_engine)
> {
> - uint8_t blk = GETFIELD(VC_ENDC_WATCH_BLOCK_ID,
> - xive->vc_regs[(VC_ENDC_WATCH0_SPEC >> 3)]);
> - uint32_t idx = GETFIELD(VC_ENDC_WATCH_INDEX,
> - xive->vc_regs[(VC_ENDC_WATCH0_SPEC >> 3)]);
> + uint8_t blk;
> + uint32_t idx;
> uint64_t endc_watch[4] = { 0 };
> - int i;
> + int i, spec_reg, data_reg;
> +
> + assert(watch_engine < ARRAY_SIZE(endc_watch));
> +
> + spec_reg = (VC_ENDC_WATCH0_SPEC + watch_engine * 0x40) >> 3;
> + data_reg = (VC_ENDC_WATCH0_DATA0 + watch_engine * 0x40) >> 3;
> + blk = GETFIELD(VC_ENDC_WATCH_BLOCK_ID, xive->vc_regs[spec_reg]);
> + idx = GETFIELD(VC_ENDC_WATCH_INDEX, xive->vc_regs[spec_reg]);
>
> if (pnv_xive2_vst_read(xive, VST_END, blk, idx, endc_watch)) {
> xive2_error(xive, "VST: no END entry %x/%x !?", blk, idx);
> }
>
> for (i = 0; i < ARRAY_SIZE(endc_watch); i++) {
> - xive->vc_regs[(VC_ENDC_WATCH0_DATA0 >> 3) + i] =
> - be64_to_cpu(endc_watch[i]);
> + xive->vc_regs[data_reg + i] = be64_to_cpu(endc_watch[i]);
> }
> }
>
> @@ -379,40 +387,50 @@ static int pnv_xive2_write_nvp(Xive2Router *xrtr, uint8_t blk, uint32_t idx,
> word_number);
> }
>
> -static int pnv_xive2_nvp_update(PnvXive2 *xive)
> +static int pnv_xive2_nvp_update(PnvXive2 *xive, uint8_t watch_engine)
> {
> - uint8_t blk = GETFIELD(PC_NXC_WATCH_BLOCK_ID,
> - xive->pc_regs[(PC_NXC_WATCH0_SPEC >> 3)]);
> - uint32_t idx = GETFIELD(PC_NXC_WATCH_INDEX,
> - xive->pc_regs[(PC_NXC_WATCH0_SPEC >> 3)]);
> - int i;
> + uint8_t blk;
> + uint32_t idx;
> + int i, spec_reg, data_reg;
> uint64_t nxc_watch[4];
>
> + if (watch_engine > 3) {
> + return -1;
> + }
Why not an assert here and below ?
Thanks,
C.
> + spec_reg = (PC_NXC_WATCH0_SPEC + watch_engine * 0x40) >> 3;
> + data_reg = (PC_NXC_WATCH0_DATA0 + watch_engine * 0x40) >> 3;
> + blk = GETFIELD(PC_NXC_WATCH_BLOCK_ID, xive->pc_regs[spec_reg]);
> + idx = GETFIELD(PC_NXC_WATCH_INDEX, xive->pc_regs[spec_reg]);
> +
> for (i = 0; i < ARRAY_SIZE(nxc_watch); i++) {
> - nxc_watch[i] =
> - cpu_to_be64(xive->pc_regs[(PC_NXC_WATCH0_DATA0 >> 3) + i]);
> + nxc_watch[i] = cpu_to_be64(xive->pc_regs[data_reg + i]);
> }
>
> return pnv_xive2_vst_write(xive, VST_NVP, blk, idx, nxc_watch,
> XIVE_VST_WORD_ALL);
> }
>
> -static void pnv_xive2_nvp_cache_load(PnvXive2 *xive)
> +static void pnv_xive2_nvp_cache_load(PnvXive2 *xive, uint8_t watch_engine)
> {
> - uint8_t blk = GETFIELD(PC_NXC_WATCH_BLOCK_ID,
> - xive->pc_regs[(PC_NXC_WATCH0_SPEC >> 3)]);
> - uint32_t idx = GETFIELD(PC_NXC_WATCH_INDEX,
> - xive->pc_regs[(PC_NXC_WATCH0_SPEC >> 3)]);
> + uint8_t blk;
> + uint32_t idx;
> uint64_t nxc_watch[4] = { 0 };
> - int i;
> + int i, spec_reg, data_reg;
> +
> + if (watch_engine > 3) {
> + return;
> + }
> + spec_reg = (PC_NXC_WATCH0_SPEC + watch_engine * 0x40) >> 3;
> + data_reg = (PC_NXC_WATCH0_DATA0 + watch_engine * 0x40) >> 3;
> + blk = GETFIELD(PC_NXC_WATCH_BLOCK_ID, xive->pc_regs[spec_reg]);
> + idx = GETFIELD(PC_NXC_WATCH_INDEX, xive->pc_regs[spec_reg]);
>
> if (pnv_xive2_vst_read(xive, VST_NVP, blk, idx, nxc_watch)) {
> xive2_error(xive, "VST: no NVP entry %x/%x !?", blk, idx);
> }
>
> for (i = 0; i < ARRAY_SIZE(nxc_watch); i++) {
> - xive->pc_regs[(PC_NXC_WATCH0_DATA0 >> 3) + i] =
> - be64_to_cpu(nxc_watch[i]);
> + xive->pc_regs[data_reg + i] = be64_to_cpu(nxc_watch[i]);
> }
> }
>
> @@ -964,12 +982,70 @@ static const MemoryRegionOps pnv_xive2_ic_cq_ops = {
> },
> };
>
> +static uint8_t pnv_xive2_cache_watch_assign(uint64_t engine_mask,
> + uint64_t *state)
> +{
> + uint8_t val = 0xFF;
> + int i;
> +
> + for (i = 3; i >= 0; i--) {
> + if (BIT(i) & engine_mask) {
> + if (!(BIT(i) & *state)) {
> + *state |= BIT(i);
> + val = 3 - i;
> + break;
> + }
> + }
> + }
> + return val;
> +}
> +
> +static void pnv_xive2_cache_watch_release(uint64_t *state, uint8_t watch_engine)
> +{
> + uint8_t engine_bit = 3 - watch_engine;
> +
> + if (*state & BIT(engine_bit)) {
> + *state &= ~BIT(engine_bit);
> + }
> +}
> +
> +static uint8_t pnv_xive2_endc_cache_watch_assign(PnvXive2 *xive)
> +{
> + uint64_t engine_mask = GETFIELD(VC_ENDC_CFG_CACHE_WATCH_ASSIGN,
> + xive->vc_regs[VC_ENDC_CFG >> 3]);
> + uint64_t state = xive->vc_regs[VC_ENDC_WATCH_ASSIGN >> 3];
> + uint8_t val;
> +
> + /*
> + * We keep track of which engines are currently busy in the
> + * VC_ENDC_WATCH_ASSIGN register directly. When the firmware reads
> + * the register, we don't return its value but the ID of an engine
> + * it can use.
> + * There are 4 engines. 0xFF means no engine is available.
> + */
> + val = pnv_xive2_cache_watch_assign(engine_mask, &state);
> + if (val != 0xFF) {
> + xive->vc_regs[VC_ENDC_WATCH_ASSIGN >> 3] = state;
> + }
> + return val;
> +}
> +
> +static void pnv_xive2_endc_cache_watch_release(PnvXive2 *xive,
> + uint8_t watch_engine)
> +{
> + uint64_t state = xive->vc_regs[VC_ENDC_WATCH_ASSIGN >> 3];
> +
> + pnv_xive2_cache_watch_release(&state, watch_engine);
> + xive->vc_regs[VC_ENDC_WATCH_ASSIGN >> 3] = state;
> +}
> +
> static uint64_t pnv_xive2_ic_vc_read(void *opaque, hwaddr offset,
> unsigned size)
> {
> PnvXive2 *xive = PNV_XIVE2(opaque);
> uint64_t val = 0;
> uint32_t reg = offset >> 3;
> + uint8_t watch_engine;
>
> switch (offset) {
> /*
> @@ -1000,24 +1076,44 @@ static uint64_t pnv_xive2_ic_vc_read(void *opaque, hwaddr offset,
> val = xive->vc_regs[reg];
> break;
>
> + case VC_ENDC_WATCH_ASSIGN:
> + val = pnv_xive2_endc_cache_watch_assign(xive);
> + break;
> +
> + case VC_ENDC_CFG:
> + val = xive->vc_regs[reg];
> + break;
> +
> /*
> * END cache updates
> */
> case VC_ENDC_WATCH0_SPEC:
> + case VC_ENDC_WATCH1_SPEC:
> + case VC_ENDC_WATCH2_SPEC:
> + case VC_ENDC_WATCH3_SPEC:
> + watch_engine = (offset - VC_ENDC_WATCH0_SPEC) >> 6;
> xive->vc_regs[reg] &= ~(VC_ENDC_WATCH_FULL | VC_ENDC_WATCH_CONFLICT);
> + pnv_xive2_endc_cache_watch_release(xive, watch_engine);
> val = xive->vc_regs[reg];
> break;
>
> case VC_ENDC_WATCH0_DATA0:
> + case VC_ENDC_WATCH1_DATA0:
> + case VC_ENDC_WATCH2_DATA0:
> + case VC_ENDC_WATCH3_DATA0:
> /*
> * Load DATA registers from cache with data requested by the
> * SPEC register
> */
> - pnv_xive2_end_cache_load(xive);
> + watch_engine = (offset - VC_ENDC_WATCH0_DATA0) >> 6;
> + pnv_xive2_end_cache_load(xive, watch_engine);
> val = xive->vc_regs[reg];
> break;
>
> case VC_ENDC_WATCH0_DATA1 ... VC_ENDC_WATCH0_DATA3:
> + case VC_ENDC_WATCH1_DATA1 ... VC_ENDC_WATCH1_DATA3:
> + case VC_ENDC_WATCH2_DATA1 ... VC_ENDC_WATCH2_DATA3:
> + case VC_ENDC_WATCH3_DATA1 ... VC_ENDC_WATCH3_DATA3:
> val = xive->vc_regs[reg];
> break;
>
> @@ -1063,6 +1159,7 @@ static void pnv_xive2_ic_vc_write(void *opaque, hwaddr offset,
> {
> PnvXive2 *xive = PNV_XIVE2(opaque);
> uint32_t reg = offset >> 3;
> + uint8_t watch_engine;
>
> switch (offset) {
> /*
> @@ -1095,19 +1192,32 @@ static void pnv_xive2_ic_vc_write(void *opaque, hwaddr offset,
> /* EAS update */
> break;
>
> + case VC_ENDC_CFG:
> + break;
> +
> /*
> * END cache updates
> */
> case VC_ENDC_WATCH0_SPEC:
> + case VC_ENDC_WATCH1_SPEC:
> + case VC_ENDC_WATCH2_SPEC:
> + case VC_ENDC_WATCH3_SPEC:
> val &= ~VC_ENDC_WATCH_CONFLICT; /* HW will set this bit */
> break;
>
> case VC_ENDC_WATCH0_DATA1 ... VC_ENDC_WATCH0_DATA3:
> + case VC_ENDC_WATCH1_DATA1 ... VC_ENDC_WATCH1_DATA3:
> + case VC_ENDC_WATCH2_DATA1 ... VC_ENDC_WATCH2_DATA3:
> + case VC_ENDC_WATCH3_DATA1 ... VC_ENDC_WATCH3_DATA3:
> break;
> case VC_ENDC_WATCH0_DATA0:
> + case VC_ENDC_WATCH1_DATA0:
> + case VC_ENDC_WATCH2_DATA0:
> + case VC_ENDC_WATCH3_DATA0:
> /* writing to DATA0 triggers the cache write */
> + watch_engine = (offset - VC_ENDC_WATCH0_DATA0) >> 6;
> xive->vc_regs[reg] = val;
> - pnv_xive2_end_update(xive);
> + pnv_xive2_end_update(xive, watch_engine);
> break;
>
>
> @@ -1157,12 +1267,43 @@ static const MemoryRegionOps pnv_xive2_ic_vc_ops = {
> },
> };
>
> +static uint8_t pnv_xive2_nxc_cache_watch_assign(PnvXive2 *xive)
> +{
> + uint64_t engine_mask = GETFIELD(PC_NXC_PROC_CONFIG_WATCH_ASSIGN,
> + xive->pc_regs[PC_NXC_PROC_CONFIG >> 3]);
> + uint64_t state = xive->pc_regs[PC_NXC_WATCH_ASSIGN >> 3];
> + uint8_t val;
> +
> + /*
> + * We keep track of which engines are currently busy in the
> + * PC_NXC_WATCH_ASSIGN register directly. When the firmware reads
> + * the register, we don't return its value but the ID of an engine
> + * it can use.
> + * There are 4 engines. 0xFF means no engine is available.
> + */
> + val = pnv_xive2_cache_watch_assign(engine_mask, &state);
> + if (val != 0xFF) {
> + xive->pc_regs[PC_NXC_WATCH_ASSIGN >> 3] = state;
> + }
> + return val;
> +}
> +
> +static void pnv_xive2_nxc_cache_watch_release(PnvXive2 *xive,
> + uint8_t watch_engine)
> +{
> + uint64_t state = xive->pc_regs[PC_NXC_WATCH_ASSIGN >> 3];
> +
> + pnv_xive2_cache_watch_release(&state, watch_engine);
> + xive->pc_regs[PC_NXC_WATCH_ASSIGN >> 3] = state;
> +}
> +
> static uint64_t pnv_xive2_ic_pc_read(void *opaque, hwaddr offset,
> unsigned size)
> {
> PnvXive2 *xive = PNV_XIVE2(opaque);
> uint64_t val = -1;
> uint32_t reg = offset >> 3;
> + uint8_t watch_engine;
>
> switch (offset) {
> /*
> @@ -1173,24 +1314,44 @@ static uint64_t pnv_xive2_ic_pc_read(void *opaque, hwaddr offset,
> val = xive->pc_regs[reg];
> break;
>
> + case PC_NXC_WATCH_ASSIGN:
> + val = pnv_xive2_nxc_cache_watch_assign(xive);
> + break;
> +
> + case PC_NXC_PROC_CONFIG:
> + val = xive->pc_regs[reg];
> + break;
> +
> /*
> * cache updates
> */
> case PC_NXC_WATCH0_SPEC:
> + case PC_NXC_WATCH1_SPEC:
> + case PC_NXC_WATCH2_SPEC:
> + case PC_NXC_WATCH3_SPEC:
> + watch_engine = (offset - PC_NXC_WATCH0_SPEC) >> 6;
> xive->pc_regs[reg] &= ~(PC_NXC_WATCH_FULL | PC_NXC_WATCH_CONFLICT);
> + pnv_xive2_nxc_cache_watch_release(xive, watch_engine);
> val = xive->pc_regs[reg];
> break;
>
> case PC_NXC_WATCH0_DATA0:
> + case PC_NXC_WATCH1_DATA0:
> + case PC_NXC_WATCH2_DATA0:
> + case PC_NXC_WATCH3_DATA0:
> /*
> * Load DATA registers from cache with data requested by the
> * SPEC register
> */
> - pnv_xive2_nvp_cache_load(xive);
> + watch_engine = (offset - PC_NXC_WATCH0_DATA0) >> 6;
> + pnv_xive2_nvp_cache_load(xive, watch_engine);
> val = xive->pc_regs[reg];
> break;
>
> case PC_NXC_WATCH0_DATA1 ... PC_NXC_WATCH0_DATA3:
> + case PC_NXC_WATCH1_DATA1 ... PC_NXC_WATCH1_DATA3:
> + case PC_NXC_WATCH2_DATA1 ... PC_NXC_WATCH2_DATA3:
> + case PC_NXC_WATCH3_DATA1 ... PC_NXC_WATCH3_DATA3:
> val = xive->pc_regs[reg];
> break;
>
> @@ -1219,6 +1380,7 @@ static void pnv_xive2_ic_pc_write(void *opaque, hwaddr offset,
> {
> PnvXive2 *xive = PNV_XIVE2(opaque);
> uint32_t reg = offset >> 3;
> + uint8_t watch_engine;
>
> switch (offset) {
>
> @@ -1231,19 +1393,32 @@ static void pnv_xive2_ic_pc_write(void *opaque, hwaddr offset,
> case PC_VSD_TABLE_DATA:
> break;
>
> + case PC_NXC_PROC_CONFIG:
> + break;
> +
> /*
> * cache updates
> */
> case PC_NXC_WATCH0_SPEC:
> + case PC_NXC_WATCH1_SPEC:
> + case PC_NXC_WATCH2_SPEC:
> + case PC_NXC_WATCH3_SPEC:
> val &= ~PC_NXC_WATCH_CONFLICT; /* HW will set this bit */
> break;
>
> case PC_NXC_WATCH0_DATA1 ... PC_NXC_WATCH0_DATA3:
> + case PC_NXC_WATCH1_DATA1 ... PC_NXC_WATCH1_DATA3:
> + case PC_NXC_WATCH2_DATA1 ... PC_NXC_WATCH2_DATA3:
> + case PC_NXC_WATCH3_DATA1 ... PC_NXC_WATCH3_DATA3:
> break;
> case PC_NXC_WATCH0_DATA0:
> + case PC_NXC_WATCH1_DATA0:
> + case PC_NXC_WATCH2_DATA0:
> + case PC_NXC_WATCH3_DATA0:
> /* writing to DATA0 triggers the cache write */
> + watch_engine = (offset - PC_NXC_WATCH0_DATA0) >> 6;
> xive->pc_regs[reg] = val;
> - pnv_xive2_nvp_update(xive);
> + pnv_xive2_nvp_update(xive, watch_engine);
> break;
>
> /* case PC_NXC_FLUSH_CTRL: */
> @@ -1814,6 +1989,12 @@ static void pnv_xive2_reset(void *dev)
> xive->cq_regs[CQ_XIVE_CFG >> 3] |=
> SETFIELD(CQ_XIVE_CFG_HYP_HARD_BLOCK_ID, 0ull, xive->chip->chip_id);
>
> + /* VC and PC cache watch assign mechanism */
> + xive->vc_regs[VC_ENDC_CFG >> 3] =
> + SETFIELD(VC_ENDC_CFG_CACHE_WATCH_ASSIGN, 0ull, 0b0111);
> + xive->pc_regs[PC_NXC_PROC_CONFIG >> 3] =
> + SETFIELD(PC_NXC_PROC_CONFIG_WATCH_ASSIGN, 0ull, 0b0111);
> +
> /* Set default page size to 64k */
> xive->ic_shift = xive->esb_shift = xive->end_shift = 16;
> xive->nvc_shift = xive->nvpg_shift = xive->tm_shift = 16;
^ permalink raw reply [flat|nested] 23+ messages in thread* Re: [PATCH v2 1/9] pnv/xive2: XIVE2 Cache Watch, Cache Flush and Sync Injection support
2024-07-16 7:28 ` Cédric Le Goater
@ 2024-07-16 15:32 ` Mike Kowal
2024-07-16 15:55 ` Mike Kowal
0 siblings, 1 reply; 23+ messages in thread
From: Mike Kowal @ 2024-07-16 15:32 UTC (permalink / raw)
To: Cédric Le Goater, Michael Kowal, qemu-devel
Cc: qemu-ppc, fbarrat, npiggin, milesg
On 7/16/2024 2:28 AM, Cédric Le Goater wrote:
> On 7/15/24 20:33, Michael Kowal wrote:
>> From: Frederic Barrat <fbarrat@linux.ibm.com>
>>
>> XIVE offers a 'cache watch facility', which allows software to
>> read/update
>> a potentially cached table entry with no software lock. There's one such
>> facility in the Virtualization Controller (VC) to update the ESB and END
>> entries and one in the Presentation Controller (PC) to update the
>> NVP/NVG/NVC entries.
>>
>> Each facility has 4 cache watch engines to control the updates and
>> firmware can request an available engine by querying the hardware
>> 'watch_assign' register of the VC or PC. The engine is then reserved and
>> is released after the data is updated by reading the 'watch_spec'
>> register
>> (which also allows to check for a conflict during the update).
>> If no engine is available, the special value 0xFF is returned and
>> firmware is expected to repeat the request until an engine becomes
>> available.
>>
>> Signed-off-by: Frederic Barrat <fbarrat@linux.ibm.com>
>> Signed-off-by: Michael Kowal <kowal@linux.vnet.ibm.com>
>> ---
>> hw/intc/pnv_xive2_regs.h | 90 ++++++++++++++
>> hw/intc/pnv_xive2.c | 253 +++++++++++++++++++++++++++++++++------
>> 2 files changed, 307 insertions(+), 36 deletions(-)
>>
>> diff --git a/hw/intc/pnv_xive2_regs.h b/hw/intc/pnv_xive2_regs.h
>> index 7165dc8704..f8e4a677c6 100644
>> --- a/hw/intc/pnv_xive2_regs.h
>> +++ b/hw/intc/pnv_xive2_regs.h
>> @@ -283,6 +283,15 @@
>> #define VC_ENDC_SYNC_QUEUE_HARD PPC_BIT(6)
>> #define VC_QUEUE_COUNT 7
>> +/* ENDC cache watch assign */
>> +#define X_VC_ENDC_WATCH_ASSIGN 0x186
>> +#define VC_ENDC_WATCH_ASSIGN 0x430
>> +
>> +/* ENDC configuration register */
>> +#define X_VC_ENDC_CFG 0x188
>> +#define VC_ENDC_CFG 0x440
>> +#define VC_ENDC_CFG_CACHE_WATCH_ASSIGN PPC_BITMASK(32, 35)
>> +
>> /* ENDC cache watch specification 0 */
>> #define X_VC_ENDC_WATCH0_SPEC 0x1A0
>> #define VC_ENDC_WATCH0_SPEC 0x500
>> @@ -302,6 +311,42 @@
>> #define VC_ENDC_WATCH0_DATA2 0x530
>> #define VC_ENDC_WATCH0_DATA3 0x538
>> +/* ENDC cache watch 1 */
>> +#define X_VC_ENDC_WATCH1_SPEC 0x1A8
>> +#define VC_ENDC_WATCH1_SPEC 0x540
>> +#define X_VC_ENDC_WATCH1_DATA0 0x1AC
>> +#define X_VC_ENDC_WATCH1_DATA1 0x1AD
>> +#define X_VC_ENDC_WATCH1_DATA2 0x1AE
>> +#define X_VC_ENDC_WATCH1_DATA3 0x1AF
>> +#define VC_ENDC_WATCH1_DATA0 0x560
>> +#define VC_ENDC_WATCH1_DATA1 0x568
>> +#define VC_ENDC_WATCH1_DATA2 0x570
>> +#define VC_ENDC_WATCH1_DATA3 0x578
>> +
>> +/* ENDC cache watch 2 */
>> +#define X_VC_ENDC_WATCH2_SPEC 0x1B0
>> +#define VC_ENDC_WATCH2_SPEC 0x580
>> +#define X_VC_ENDC_WATCH2_DATA0 0x1B4
>> +#define X_VC_ENDC_WATCH2_DATA1 0x1B5
>> +#define X_VC_ENDC_WATCH2_DATA2 0x1B6
>> +#define X_VC_ENDC_WATCH2_DATA3 0x1B7
>> +#define VC_ENDC_WATCH2_DATA0 0x5A0
>> +#define VC_ENDC_WATCH2_DATA1 0x5A8
>> +#define VC_ENDC_WATCH2_DATA2 0x5B0
>> +#define VC_ENDC_WATCH2_DATA3 0x5B8
>> +
>> +/* ENDC cache watch 3 */
>> +#define X_VC_ENDC_WATCH3_SPEC 0x1B8
>> +#define VC_ENDC_WATCH3_SPEC 0x5C0
>> +#define X_VC_ENDC_WATCH3_DATA0 0x1BC
>> +#define X_VC_ENDC_WATCH3_DATA1 0x1BD
>> +#define X_VC_ENDC_WATCH3_DATA2 0x1BE
>> +#define X_VC_ENDC_WATCH3_DATA3 0x1BF
>> +#define VC_ENDC_WATCH3_DATA0 0x5E0
>> +#define VC_ENDC_WATCH3_DATA1 0x5E8
>> +#define VC_ENDC_WATCH3_DATA2 0x5F0
>> +#define VC_ENDC_WATCH3_DATA3 0x5F8
>> +
>> /*
>> * PC LSB1
>> */
>> @@ -358,6 +403,15 @@
>> #define PC_NXC_FLUSH_POLL_BLOCK_ID_MASK PPC_BITMASK(36, 39)
>> #define PC_NXC_FLUSH_POLL_OFFSET_MASK PPC_BITMASK(40, 63) /*
>> 24-bit */
>> +/* NxC Cache watch assign */
>> +#define X_PC_NXC_WATCH_ASSIGN 0x286
>> +#define PC_NXC_WATCH_ASSIGN 0x430
>> +
>> +/* NxC Proc config */
>> +#define X_PC_NXC_PROC_CONFIG 0x28A
>> +#define PC_NXC_PROC_CONFIG 0x450
>> +#define PC_NXC_PROC_CONFIG_WATCH_ASSIGN PPC_BITMASK(0, 3)
>> +
>> /* NxC Cache Watch 0 Specification */
>> #define X_PC_NXC_WATCH0_SPEC 0x2A0
>> #define PC_NXC_WATCH0_SPEC 0x500
>> @@ -381,6 +435,42 @@
>> #define PC_NXC_WATCH0_DATA2 0x530
>> #define PC_NXC_WATCH0_DATA3 0x538
>> +/* NxC Cache Watch 1 */
>> +#define X_PC_NXC_WATCH1_SPEC 0x2A8
>> +#define PC_NXC_WATCH1_SPEC 0x540
>> +#define X_PC_NXC_WATCH1_DATA0 0x2AC
>> +#define X_PC_NXC_WATCH1_DATA1 0x2AD
>> +#define X_PC_NXC_WATCH1_DATA2 0x2AE
>> +#define X_PC_NXC_WATCH1_DATA3 0x2AF
>> +#define PC_NXC_WATCH1_DATA0 0x560
>> +#define PC_NXC_WATCH1_DATA1 0x568
>> +#define PC_NXC_WATCH1_DATA2 0x570
>> +#define PC_NXC_WATCH1_DATA3 0x578
>> +
>> +/* NxC Cache Watch 2 */
>> +#define X_PC_NXC_WATCH2_SPEC 0x2B0
>> +#define PC_NXC_WATCH2_SPEC 0x580
>> +#define X_PC_NXC_WATCH2_DATA0 0x2B4
>> +#define X_PC_NXC_WATCH2_DATA1 0x2B5
>> +#define X_PC_NXC_WATCH2_DATA2 0x2B6
>> +#define X_PC_NXC_WATCH2_DATA3 0x2B7
>> +#define PC_NXC_WATCH2_DATA0 0x5A0
>> +#define PC_NXC_WATCH2_DATA1 0x5A8
>> +#define PC_NXC_WATCH2_DATA2 0x5B0
>> +#define PC_NXC_WATCH2_DATA3 0x5B8
>> +
>> +/* NxC Cache Watch 3 */
>> +#define X_PC_NXC_WATCH3_SPEC 0x2B8
>> +#define PC_NXC_WATCH3_SPEC 0x5C0
>> +#define X_PC_NXC_WATCH3_DATA0 0x2BC
>> +#define X_PC_NXC_WATCH3_DATA1 0x2BD
>> +#define X_PC_NXC_WATCH3_DATA2 0x2BE
>> +#define X_PC_NXC_WATCH3_DATA3 0x2BF
>> +#define PC_NXC_WATCH3_DATA0 0x5E0
>> +#define PC_NXC_WATCH3_DATA1 0x5E8
>> +#define PC_NXC_WATCH3_DATA2 0x5F0
>> +#define PC_NXC_WATCH3_DATA3 0x5F8
>> +
>> /*
>> * TCTXT Registers
>> */
>> diff --git a/hw/intc/pnv_xive2.c b/hw/intc/pnv_xive2.c
>> index 2fb4fa29d4..f6a735cca5 100644
>> --- a/hw/intc/pnv_xive2.c
>> +++ b/hw/intc/pnv_xive2.c
>> @@ -329,40 +329,48 @@ static int pnv_xive2_write_end(Xive2Router
>> *xrtr, uint8_t blk, uint32_t idx,
>> word_number);
>> }
>> -static int pnv_xive2_end_update(PnvXive2 *xive)
>> +static int pnv_xive2_end_update(PnvXive2 *xive, uint8_t watch_engine)
>> {
>> - uint8_t blk = GETFIELD(VC_ENDC_WATCH_BLOCK_ID,
>> - xive->vc_regs[(VC_ENDC_WATCH0_SPEC >> 3)]);
>> - uint32_t idx = GETFIELD(VC_ENDC_WATCH_INDEX,
>> - xive->vc_regs[(VC_ENDC_WATCH0_SPEC >> 3)]);
>> - int i;
>> + uint8_t blk;
>> + uint32_t idx;
>> + int i, spec_reg, data_reg;
>> uint64_t endc_watch[4];
>> + assert(watch_engine < ARRAY_SIZE(endc_watch));
>> +
>> + spec_reg = (VC_ENDC_WATCH0_SPEC + watch_engine * 0x40) >> 3;
>> + data_reg = (VC_ENDC_WATCH0_DATA0 + watch_engine * 0x40) >> 3;
>> + blk = GETFIELD(VC_ENDC_WATCH_BLOCK_ID, xive->vc_regs[spec_reg]);
>> + idx = GETFIELD(VC_ENDC_WATCH_INDEX, xive->vc_regs[spec_reg]);
>> +
>> for (i = 0; i < ARRAY_SIZE(endc_watch); i++) {
>> - endc_watch[i] =
>> - cpu_to_be64(xive->vc_regs[(VC_ENDC_WATCH0_DATA0 >> 3) +
>> i]);
>> + endc_watch[i] = cpu_to_be64(xive->vc_regs[data_reg + i]);
>> }
>> return pnv_xive2_vst_write(xive, VST_END, blk, idx, endc_watch,
>> XIVE_VST_WORD_ALL);
>> }
>> -static void pnv_xive2_end_cache_load(PnvXive2 *xive)
>> +static void pnv_xive2_end_cache_load(PnvXive2 *xive, uint8_t
>> watch_engine)
>> {
>> - uint8_t blk = GETFIELD(VC_ENDC_WATCH_BLOCK_ID,
>> - xive->vc_regs[(VC_ENDC_WATCH0_SPEC >> 3)]);
>> - uint32_t idx = GETFIELD(VC_ENDC_WATCH_INDEX,
>> - xive->vc_regs[(VC_ENDC_WATCH0_SPEC >> 3)]);
>> + uint8_t blk;
>> + uint32_t idx;
>> uint64_t endc_watch[4] = { 0 };
>> - int i;
>> + int i, spec_reg, data_reg;
>> +
>> + assert(watch_engine < ARRAY_SIZE(endc_watch));
>> +
>> + spec_reg = (VC_ENDC_WATCH0_SPEC + watch_engine * 0x40) >> 3;
>> + data_reg = (VC_ENDC_WATCH0_DATA0 + watch_engine * 0x40) >> 3;
>> + blk = GETFIELD(VC_ENDC_WATCH_BLOCK_ID, xive->vc_regs[spec_reg]);
>> + idx = GETFIELD(VC_ENDC_WATCH_INDEX, xive->vc_regs[spec_reg]);
>> if (pnv_xive2_vst_read(xive, VST_END, blk, idx, endc_watch)) {
>> xive2_error(xive, "VST: no END entry %x/%x !?", blk, idx);
>> }
>> for (i = 0; i < ARRAY_SIZE(endc_watch); i++) {
>> - xive->vc_regs[(VC_ENDC_WATCH0_DATA0 >> 3) + i] =
>> - be64_to_cpu(endc_watch[i]);
>> + xive->vc_regs[data_reg + i] = be64_to_cpu(endc_watch[i]);
>> }
>> }
>> @@ -379,40 +387,50 @@ static int pnv_xive2_write_nvp(Xive2Router
>> *xrtr, uint8_t blk, uint32_t idx,
>> word_number);
>> }
>> -static int pnv_xive2_nvp_update(PnvXive2 *xive)
>> +static int pnv_xive2_nvp_update(PnvXive2 *xive, uint8_t watch_engine)
>> {
>> - uint8_t blk = GETFIELD(PC_NXC_WATCH_BLOCK_ID,
>> - xive->pc_regs[(PC_NXC_WATCH0_SPEC >> 3)]);
>> - uint32_t idx = GETFIELD(PC_NXC_WATCH_INDEX,
>> - xive->pc_regs[(PC_NXC_WATCH0_SPEC >> 3)]);
>> - int i;
>> + uint8_t blk;
>> + uint32_t idx;
>> + int i, spec_reg, data_reg;
>> uint64_t nxc_watch[4];
>> + if (watch_engine > 3) {
>> + return -1;
>> + }
>
> Why not an assert here and below ?
>
>
> Thanks,
>
> C.
>
Hmmm, I added assert()s... Not sure what happened, maybe applying a
different patch overwrote it. I will verify changes and add assert()s
back in.
>
>
>
>> + spec_reg = (PC_NXC_WATCH0_SPEC + watch_engine * 0x40) >> 3;
>> + data_reg = (PC_NXC_WATCH0_DATA0 + watch_engine * 0x40) >> 3;
>> + blk = GETFIELD(PC_NXC_WATCH_BLOCK_ID, xive->pc_regs[spec_reg]);
>> + idx = GETFIELD(PC_NXC_WATCH_INDEX, xive->pc_regs[spec_reg]);
>> +
>> for (i = 0; i < ARRAY_SIZE(nxc_watch); i++) {
>> - nxc_watch[i] =
>> - cpu_to_be64(xive->pc_regs[(PC_NXC_WATCH0_DATA0 >> 3) + i]);
>> + nxc_watch[i] = cpu_to_be64(xive->pc_regs[data_reg + i]);
>> }
>> return pnv_xive2_vst_write(xive, VST_NVP, blk, idx, nxc_watch,
>> XIVE_VST_WORD_ALL);
>> }
>> -static void pnv_xive2_nvp_cache_load(PnvXive2 *xive)
>> +static void pnv_xive2_nvp_cache_load(PnvXive2 *xive, uint8_t
>> watch_engine)
>> {
>> - uint8_t blk = GETFIELD(PC_NXC_WATCH_BLOCK_ID,
>> - xive->pc_regs[(PC_NXC_WATCH0_SPEC >> 3)]);
>> - uint32_t idx = GETFIELD(PC_NXC_WATCH_INDEX,
>> - xive->pc_regs[(PC_NXC_WATCH0_SPEC >> 3)]);
>> + uint8_t blk;
>> + uint32_t idx;
>> uint64_t nxc_watch[4] = { 0 };
>> - int i;
>> + int i, spec_reg, data_reg;
>> +
>> + if (watch_engine > 3) {
>> + return;
>> + }
>> + spec_reg = (PC_NXC_WATCH0_SPEC + watch_engine * 0x40) >> 3;
>> + data_reg = (PC_NXC_WATCH0_DATA0 + watch_engine * 0x40) >> 3;
>> + blk = GETFIELD(PC_NXC_WATCH_BLOCK_ID, xive->pc_regs[spec_reg]);
>> + idx = GETFIELD(PC_NXC_WATCH_INDEX, xive->pc_regs[spec_reg]);
>> if (pnv_xive2_vst_read(xive, VST_NVP, blk, idx, nxc_watch)) {
>> xive2_error(xive, "VST: no NVP entry %x/%x !?", blk, idx);
>> }
>> for (i = 0; i < ARRAY_SIZE(nxc_watch); i++) {
>> - xive->pc_regs[(PC_NXC_WATCH0_DATA0 >> 3) + i] =
>> - be64_to_cpu(nxc_watch[i]);
>> + xive->pc_regs[data_reg + i] = be64_to_cpu(nxc_watch[i]);
>> }
>> }
>> @@ -964,12 +982,70 @@ static const MemoryRegionOps
>> pnv_xive2_ic_cq_ops = {
>> },
>> };
>> +static uint8_t pnv_xive2_cache_watch_assign(uint64_t engine_mask,
>> + uint64_t *state)
>> +{
>> + uint8_t val = 0xFF;
>> + int i;
>> +
>> + for (i = 3; i >= 0; i--) {
>> + if (BIT(i) & engine_mask) {
>> + if (!(BIT(i) & *state)) {
>> + *state |= BIT(i);
>> + val = 3 - i;
>> + break;
>> + }
>> + }
>> + }
>> + return val;
>> +}
>> +
>> +static void pnv_xive2_cache_watch_release(uint64_t *state, uint8_t
>> watch_engine)
>> +{
>> + uint8_t engine_bit = 3 - watch_engine;
>> +
>> + if (*state & BIT(engine_bit)) {
>> + *state &= ~BIT(engine_bit);
>> + }
>> +}
>> +
>> +static uint8_t pnv_xive2_endc_cache_watch_assign(PnvXive2 *xive)
>> +{
>> + uint64_t engine_mask = GETFIELD(VC_ENDC_CFG_CACHE_WATCH_ASSIGN,
>> + xive->vc_regs[VC_ENDC_CFG >> 3]);
>> + uint64_t state = xive->vc_regs[VC_ENDC_WATCH_ASSIGN >> 3];
>> + uint8_t val;
>> +
>> + /*
>> + * We keep track of which engines are currently busy in the
>> + * VC_ENDC_WATCH_ASSIGN register directly. When the firmware reads
>> + * the register, we don't return its value but the ID of an engine
>> + * it can use.
>> + * There are 4 engines. 0xFF means no engine is available.
>> + */
>> + val = pnv_xive2_cache_watch_assign(engine_mask, &state);
>> + if (val != 0xFF) {
>> + xive->vc_regs[VC_ENDC_WATCH_ASSIGN >> 3] = state;
>> + }
>> + return val;
>> +}
>> +
>> +static void pnv_xive2_endc_cache_watch_release(PnvXive2 *xive,
>> + uint8_t watch_engine)
>> +{
>> + uint64_t state = xive->vc_regs[VC_ENDC_WATCH_ASSIGN >> 3];
>> +
>> + pnv_xive2_cache_watch_release(&state, watch_engine);
>> + xive->vc_regs[VC_ENDC_WATCH_ASSIGN >> 3] = state;
>> +}
>> +
>> static uint64_t pnv_xive2_ic_vc_read(void *opaque, hwaddr offset,
>> unsigned size)
>> {
>> PnvXive2 *xive = PNV_XIVE2(opaque);
>> uint64_t val = 0;
>> uint32_t reg = offset >> 3;
>> + uint8_t watch_engine;
>> switch (offset) {
>> /*
>> @@ -1000,24 +1076,44 @@ static uint64_t pnv_xive2_ic_vc_read(void
>> *opaque, hwaddr offset,
>> val = xive->vc_regs[reg];
>> break;
>> + case VC_ENDC_WATCH_ASSIGN:
>> + val = pnv_xive2_endc_cache_watch_assign(xive);
>> + break;
>> +
>> + case VC_ENDC_CFG:
>> + val = xive->vc_regs[reg];
>> + break;
>> +
>> /*
>> * END cache updates
>> */
>> case VC_ENDC_WATCH0_SPEC:
>> + case VC_ENDC_WATCH1_SPEC:
>> + case VC_ENDC_WATCH2_SPEC:
>> + case VC_ENDC_WATCH3_SPEC:
>> + watch_engine = (offset - VC_ENDC_WATCH0_SPEC) >> 6;
>> xive->vc_regs[reg] &= ~(VC_ENDC_WATCH_FULL |
>> VC_ENDC_WATCH_CONFLICT);
>> + pnv_xive2_endc_cache_watch_release(xive, watch_engine);
>> val = xive->vc_regs[reg];
>> break;
>> case VC_ENDC_WATCH0_DATA0:
>> + case VC_ENDC_WATCH1_DATA0:
>> + case VC_ENDC_WATCH2_DATA0:
>> + case VC_ENDC_WATCH3_DATA0:
>> /*
>> * Load DATA registers from cache with data requested by the
>> * SPEC register
>> */
>> - pnv_xive2_end_cache_load(xive);
>> + watch_engine = (offset - VC_ENDC_WATCH0_DATA0) >> 6;
>> + pnv_xive2_end_cache_load(xive, watch_engine);
>> val = xive->vc_regs[reg];
>> break;
>> case VC_ENDC_WATCH0_DATA1 ... VC_ENDC_WATCH0_DATA3:
>> + case VC_ENDC_WATCH1_DATA1 ... VC_ENDC_WATCH1_DATA3:
>> + case VC_ENDC_WATCH2_DATA1 ... VC_ENDC_WATCH2_DATA3:
>> + case VC_ENDC_WATCH3_DATA1 ... VC_ENDC_WATCH3_DATA3:
>> val = xive->vc_regs[reg];
>> break;
>> @@ -1063,6 +1159,7 @@ static void pnv_xive2_ic_vc_write(void
>> *opaque, hwaddr offset,
>> {
>> PnvXive2 *xive = PNV_XIVE2(opaque);
>> uint32_t reg = offset >> 3;
>> + uint8_t watch_engine;
>> switch (offset) {
>> /*
>> @@ -1095,19 +1192,32 @@ static void pnv_xive2_ic_vc_write(void
>> *opaque, hwaddr offset,
>> /* EAS update */
>> break;
>> + case VC_ENDC_CFG:
>> + break;
>> +
>> /*
>> * END cache updates
>> */
>> case VC_ENDC_WATCH0_SPEC:
>> + case VC_ENDC_WATCH1_SPEC:
>> + case VC_ENDC_WATCH2_SPEC:
>> + case VC_ENDC_WATCH3_SPEC:
>> val &= ~VC_ENDC_WATCH_CONFLICT; /* HW will set this bit */
>> break;
>> case VC_ENDC_WATCH0_DATA1 ... VC_ENDC_WATCH0_DATA3:
>> + case VC_ENDC_WATCH1_DATA1 ... VC_ENDC_WATCH1_DATA3:
>> + case VC_ENDC_WATCH2_DATA1 ... VC_ENDC_WATCH2_DATA3:
>> + case VC_ENDC_WATCH3_DATA1 ... VC_ENDC_WATCH3_DATA3:
>> break;
>> case VC_ENDC_WATCH0_DATA0:
>> + case VC_ENDC_WATCH1_DATA0:
>> + case VC_ENDC_WATCH2_DATA0:
>> + case VC_ENDC_WATCH3_DATA0:
>> /* writing to DATA0 triggers the cache write */
>> + watch_engine = (offset - VC_ENDC_WATCH0_DATA0) >> 6;
>> xive->vc_regs[reg] = val;
>> - pnv_xive2_end_update(xive);
>> + pnv_xive2_end_update(xive, watch_engine);
>> break;
>> @@ -1157,12 +1267,43 @@ static const MemoryRegionOps
>> pnv_xive2_ic_vc_ops = {
>> },
>> };
>> +static uint8_t pnv_xive2_nxc_cache_watch_assign(PnvXive2 *xive)
>> +{
>> + uint64_t engine_mask = GETFIELD(PC_NXC_PROC_CONFIG_WATCH_ASSIGN,
>> + xive->pc_regs[PC_NXC_PROC_CONFIG >> 3]);
>> + uint64_t state = xive->pc_regs[PC_NXC_WATCH_ASSIGN >> 3];
>> + uint8_t val;
>> +
>> + /*
>> + * We keep track of which engines are currently busy in the
>> + * PC_NXC_WATCH_ASSIGN register directly. When the firmware reads
>> + * the register, we don't return its value but the ID of an engine
>> + * it can use.
>> + * There are 4 engines. 0xFF means no engine is available.
>> + */
>> + val = pnv_xive2_cache_watch_assign(engine_mask, &state);
>> + if (val != 0xFF) {
>> + xive->pc_regs[PC_NXC_WATCH_ASSIGN >> 3] = state;
>> + }
>> + return val;
>> +}
>> +
>> +static void pnv_xive2_nxc_cache_watch_release(PnvXive2 *xive,
>> + uint8_t watch_engine)
>> +{
>> + uint64_t state = xive->pc_regs[PC_NXC_WATCH_ASSIGN >> 3];
>> +
>> + pnv_xive2_cache_watch_release(&state, watch_engine);
>> + xive->pc_regs[PC_NXC_WATCH_ASSIGN >> 3] = state;
>> +}
>> +
>> static uint64_t pnv_xive2_ic_pc_read(void *opaque, hwaddr offset,
>> unsigned size)
>> {
>> PnvXive2 *xive = PNV_XIVE2(opaque);
>> uint64_t val = -1;
>> uint32_t reg = offset >> 3;
>> + uint8_t watch_engine;
>> switch (offset) {
>> /*
>> @@ -1173,24 +1314,44 @@ static uint64_t pnv_xive2_ic_pc_read(void
>> *opaque, hwaddr offset,
>> val = xive->pc_regs[reg];
>> break;
>> + case PC_NXC_WATCH_ASSIGN:
>> + val = pnv_xive2_nxc_cache_watch_assign(xive);
>> + break;
>> +
>> + case PC_NXC_PROC_CONFIG:
>> + val = xive->pc_regs[reg];
>> + break;
>> +
>> /*
>> * cache updates
>> */
>> case PC_NXC_WATCH0_SPEC:
>> + case PC_NXC_WATCH1_SPEC:
>> + case PC_NXC_WATCH2_SPEC:
>> + case PC_NXC_WATCH3_SPEC:
>> + watch_engine = (offset - PC_NXC_WATCH0_SPEC) >> 6;
>> xive->pc_regs[reg] &= ~(PC_NXC_WATCH_FULL |
>> PC_NXC_WATCH_CONFLICT);
>> + pnv_xive2_nxc_cache_watch_release(xive, watch_engine);
>> val = xive->pc_regs[reg];
>> break;
>> case PC_NXC_WATCH0_DATA0:
>> + case PC_NXC_WATCH1_DATA0:
>> + case PC_NXC_WATCH2_DATA0:
>> + case PC_NXC_WATCH3_DATA0:
>> /*
>> * Load DATA registers from cache with data requested by the
>> * SPEC register
>> */
>> - pnv_xive2_nvp_cache_load(xive);
>> + watch_engine = (offset - PC_NXC_WATCH0_DATA0) >> 6;
>> + pnv_xive2_nvp_cache_load(xive, watch_engine);
>> val = xive->pc_regs[reg];
>> break;
>> case PC_NXC_WATCH0_DATA1 ... PC_NXC_WATCH0_DATA3:
>> + case PC_NXC_WATCH1_DATA1 ... PC_NXC_WATCH1_DATA3:
>> + case PC_NXC_WATCH2_DATA1 ... PC_NXC_WATCH2_DATA3:
>> + case PC_NXC_WATCH3_DATA1 ... PC_NXC_WATCH3_DATA3:
>> val = xive->pc_regs[reg];
>> break;
>> @@ -1219,6 +1380,7 @@ static void pnv_xive2_ic_pc_write(void
>> *opaque, hwaddr offset,
>> {
>> PnvXive2 *xive = PNV_XIVE2(opaque);
>> uint32_t reg = offset >> 3;
>> + uint8_t watch_engine;
>> switch (offset) {
>> @@ -1231,19 +1393,32 @@ static void pnv_xive2_ic_pc_write(void
>> *opaque, hwaddr offset,
>> case PC_VSD_TABLE_DATA:
>> break;
>> + case PC_NXC_PROC_CONFIG:
>> + break;
>> +
>> /*
>> * cache updates
>> */
>> case PC_NXC_WATCH0_SPEC:
>> + case PC_NXC_WATCH1_SPEC:
>> + case PC_NXC_WATCH2_SPEC:
>> + case PC_NXC_WATCH3_SPEC:
>> val &= ~PC_NXC_WATCH_CONFLICT; /* HW will set this bit */
>> break;
>> case PC_NXC_WATCH0_DATA1 ... PC_NXC_WATCH0_DATA3:
>> + case PC_NXC_WATCH1_DATA1 ... PC_NXC_WATCH1_DATA3:
>> + case PC_NXC_WATCH2_DATA1 ... PC_NXC_WATCH2_DATA3:
>> + case PC_NXC_WATCH3_DATA1 ... PC_NXC_WATCH3_DATA3:
>> break;
>> case PC_NXC_WATCH0_DATA0:
>> + case PC_NXC_WATCH1_DATA0:
>> + case PC_NXC_WATCH2_DATA0:
>> + case PC_NXC_WATCH3_DATA0:
>> /* writing to DATA0 triggers the cache write */
>> + watch_engine = (offset - PC_NXC_WATCH0_DATA0) >> 6;
>> xive->pc_regs[reg] = val;
>> - pnv_xive2_nvp_update(xive);
>> + pnv_xive2_nvp_update(xive, watch_engine);
>> break;
>> /* case PC_NXC_FLUSH_CTRL: */
>> @@ -1814,6 +1989,12 @@ static void pnv_xive2_reset(void *dev)
>> xive->cq_regs[CQ_XIVE_CFG >> 3] |=
>> SETFIELD(CQ_XIVE_CFG_HYP_HARD_BLOCK_ID, 0ull,
>> xive->chip->chip_id);
>> + /* VC and PC cache watch assign mechanism */
>> + xive->vc_regs[VC_ENDC_CFG >> 3] =
>> + SETFIELD(VC_ENDC_CFG_CACHE_WATCH_ASSIGN, 0ull, 0b0111);
>> + xive->pc_regs[PC_NXC_PROC_CONFIG >> 3] =
>> + SETFIELD(PC_NXC_PROC_CONFIG_WATCH_ASSIGN, 0ull, 0b0111);
>> +
>> /* Set default page size to 64k */
>> xive->ic_shift = xive->esb_shift = xive->end_shift = 16;
>> xive->nvc_shift = xive->nvpg_shift = xive->tm_shift = 16;
>
^ permalink raw reply [flat|nested] 23+ messages in thread* Re: [PATCH v2 1/9] pnv/xive2: XIVE2 Cache Watch, Cache Flush and Sync Injection support
2024-07-16 15:32 ` Mike Kowal
@ 2024-07-16 15:55 ` Mike Kowal
0 siblings, 0 replies; 23+ messages in thread
From: Mike Kowal @ 2024-07-16 15:55 UTC (permalink / raw)
To: Cédric Le Goater, Michael Kowal, qemu-devel
Cc: qemu-ppc, fbarrat, npiggin, milesg
On 7/16/2024 10:32 AM, Mike Kowal wrote:
>
> On 7/16/2024 2:28 AM, Cédric Le Goater wrote:
>> On 7/15/24 20:33, Michael Kowal wrote:
>>> From: Frederic Barrat <fbarrat@linux.ibm.com>
>>>
>>> XIVE offers a 'cache watch facility', which allows software to
>>> read/update
>>> a potentially cached table entry with no software lock. There's one
>>> such
>>> facility in the Virtualization Controller (VC) to update the ESB and
>>> END
>>> entries and one in the Presentation Controller (PC) to update the
>>> NVP/NVG/NVC entries.
>>>
>>> Each facility has 4 cache watch engines to control the updates and
>>> firmware can request an available engine by querying the hardware
>>> 'watch_assign' register of the VC or PC. The engine is then reserved
>>> and
>>> is released after the data is updated by reading the 'watch_spec'
>>> register
>>> (which also allows to check for a conflict during the update).
>>> If no engine is available, the special value 0xFF is returned and
>>> firmware is expected to repeat the request until an engine becomes
>>> available.
>>>
>>> Signed-off-by: Frederic Barrat <fbarrat@linux.ibm.com>
>>> Signed-off-by: Michael Kowal <kowal@linux.vnet.ibm.com>
>>> ---
>>> hw/intc/pnv_xive2_regs.h | 90 ++++++++++++++
>>> hw/intc/pnv_xive2.c | 253
>>> +++++++++++++++++++++++++++++++++------
>>> 2 files changed, 307 insertions(+), 36 deletions(-)
>>>
>>> diff --git a/hw/intc/pnv_xive2_regs.h b/hw/intc/pnv_xive2_regs.h
>>> index 7165dc8704..f8e4a677c6 100644
>>> --- a/hw/intc/pnv_xive2_regs.h
>>> +++ b/hw/intc/pnv_xive2_regs.h
>>> @@ -283,6 +283,15 @@
>>> #define VC_ENDC_SYNC_QUEUE_HARD PPC_BIT(6)
>>> #define VC_QUEUE_COUNT 7
>>> +/* ENDC cache watch assign */
>>> +#define X_VC_ENDC_WATCH_ASSIGN 0x186
>>> +#define VC_ENDC_WATCH_ASSIGN 0x430
>>> +
>>> +/* ENDC configuration register */
>>> +#define X_VC_ENDC_CFG 0x188
>>> +#define VC_ENDC_CFG 0x440
>>> +#define VC_ENDC_CFG_CACHE_WATCH_ASSIGN PPC_BITMASK(32, 35)
>>> +
>>> /* ENDC cache watch specification 0 */
>>> #define X_VC_ENDC_WATCH0_SPEC 0x1A0
>>> #define VC_ENDC_WATCH0_SPEC 0x500
>>> @@ -302,6 +311,42 @@
>>> #define VC_ENDC_WATCH0_DATA2 0x530
>>> #define VC_ENDC_WATCH0_DATA3 0x538
>>> +/* ENDC cache watch 1 */
>>> +#define X_VC_ENDC_WATCH1_SPEC 0x1A8
>>> +#define VC_ENDC_WATCH1_SPEC 0x540
>>> +#define X_VC_ENDC_WATCH1_DATA0 0x1AC
>>> +#define X_VC_ENDC_WATCH1_DATA1 0x1AD
>>> +#define X_VC_ENDC_WATCH1_DATA2 0x1AE
>>> +#define X_VC_ENDC_WATCH1_DATA3 0x1AF
>>> +#define VC_ENDC_WATCH1_DATA0 0x560
>>> +#define VC_ENDC_WATCH1_DATA1 0x568
>>> +#define VC_ENDC_WATCH1_DATA2 0x570
>>> +#define VC_ENDC_WATCH1_DATA3 0x578
>>> +
>>> +/* ENDC cache watch 2 */
>>> +#define X_VC_ENDC_WATCH2_SPEC 0x1B0
>>> +#define VC_ENDC_WATCH2_SPEC 0x580
>>> +#define X_VC_ENDC_WATCH2_DATA0 0x1B4
>>> +#define X_VC_ENDC_WATCH2_DATA1 0x1B5
>>> +#define X_VC_ENDC_WATCH2_DATA2 0x1B6
>>> +#define X_VC_ENDC_WATCH2_DATA3 0x1B7
>>> +#define VC_ENDC_WATCH2_DATA0 0x5A0
>>> +#define VC_ENDC_WATCH2_DATA1 0x5A8
>>> +#define VC_ENDC_WATCH2_DATA2 0x5B0
>>> +#define VC_ENDC_WATCH2_DATA3 0x5B8
>>> +
>>> +/* ENDC cache watch 3 */
>>> +#define X_VC_ENDC_WATCH3_SPEC 0x1B8
>>> +#define VC_ENDC_WATCH3_SPEC 0x5C0
>>> +#define X_VC_ENDC_WATCH3_DATA0 0x1BC
>>> +#define X_VC_ENDC_WATCH3_DATA1 0x1BD
>>> +#define X_VC_ENDC_WATCH3_DATA2 0x1BE
>>> +#define X_VC_ENDC_WATCH3_DATA3 0x1BF
>>> +#define VC_ENDC_WATCH3_DATA0 0x5E0
>>> +#define VC_ENDC_WATCH3_DATA1 0x5E8
>>> +#define VC_ENDC_WATCH3_DATA2 0x5F0
>>> +#define VC_ENDC_WATCH3_DATA3 0x5F8
>>> +
>>> /*
>>> * PC LSB1
>>> */
>>> @@ -358,6 +403,15 @@
>>> #define PC_NXC_FLUSH_POLL_BLOCK_ID_MASK PPC_BITMASK(36, 39)
>>> #define PC_NXC_FLUSH_POLL_OFFSET_MASK PPC_BITMASK(40, 63) /*
>>> 24-bit */
>>> +/* NxC Cache watch assign */
>>> +#define X_PC_NXC_WATCH_ASSIGN 0x286
>>> +#define PC_NXC_WATCH_ASSIGN 0x430
>>> +
>>> +/* NxC Proc config */
>>> +#define X_PC_NXC_PROC_CONFIG 0x28A
>>> +#define PC_NXC_PROC_CONFIG 0x450
>>> +#define PC_NXC_PROC_CONFIG_WATCH_ASSIGN PPC_BITMASK(0, 3)
>>> +
>>> /* NxC Cache Watch 0 Specification */
>>> #define X_PC_NXC_WATCH0_SPEC 0x2A0
>>> #define PC_NXC_WATCH0_SPEC 0x500
>>> @@ -381,6 +435,42 @@
>>> #define PC_NXC_WATCH0_DATA2 0x530
>>> #define PC_NXC_WATCH0_DATA3 0x538
>>> +/* NxC Cache Watch 1 */
>>> +#define X_PC_NXC_WATCH1_SPEC 0x2A8
>>> +#define PC_NXC_WATCH1_SPEC 0x540
>>> +#define X_PC_NXC_WATCH1_DATA0 0x2AC
>>> +#define X_PC_NXC_WATCH1_DATA1 0x2AD
>>> +#define X_PC_NXC_WATCH1_DATA2 0x2AE
>>> +#define X_PC_NXC_WATCH1_DATA3 0x2AF
>>> +#define PC_NXC_WATCH1_DATA0 0x560
>>> +#define PC_NXC_WATCH1_DATA1 0x568
>>> +#define PC_NXC_WATCH1_DATA2 0x570
>>> +#define PC_NXC_WATCH1_DATA3 0x578
>>> +
>>> +/* NxC Cache Watch 2 */
>>> +#define X_PC_NXC_WATCH2_SPEC 0x2B0
>>> +#define PC_NXC_WATCH2_SPEC 0x580
>>> +#define X_PC_NXC_WATCH2_DATA0 0x2B4
>>> +#define X_PC_NXC_WATCH2_DATA1 0x2B5
>>> +#define X_PC_NXC_WATCH2_DATA2 0x2B6
>>> +#define X_PC_NXC_WATCH2_DATA3 0x2B7
>>> +#define PC_NXC_WATCH2_DATA0 0x5A0
>>> +#define PC_NXC_WATCH2_DATA1 0x5A8
>>> +#define PC_NXC_WATCH2_DATA2 0x5B0
>>> +#define PC_NXC_WATCH2_DATA3 0x5B8
>>> +
>>> +/* NxC Cache Watch 3 */
>>> +#define X_PC_NXC_WATCH3_SPEC 0x2B8
>>> +#define PC_NXC_WATCH3_SPEC 0x5C0
>>> +#define X_PC_NXC_WATCH3_DATA0 0x2BC
>>> +#define X_PC_NXC_WATCH3_DATA1 0x2BD
>>> +#define X_PC_NXC_WATCH3_DATA2 0x2BE
>>> +#define X_PC_NXC_WATCH3_DATA3 0x2BF
>>> +#define PC_NXC_WATCH3_DATA0 0x5E0
>>> +#define PC_NXC_WATCH3_DATA1 0x5E8
>>> +#define PC_NXC_WATCH3_DATA2 0x5F0
>>> +#define PC_NXC_WATCH3_DATA3 0x5F8
>>> +
>>> /*
>>> * TCTXT Registers
>>> */
>>> diff --git a/hw/intc/pnv_xive2.c b/hw/intc/pnv_xive2.c
>>> index 2fb4fa29d4..f6a735cca5 100644
>>> --- a/hw/intc/pnv_xive2.c
>>> +++ b/hw/intc/pnv_xive2.c
>>> @@ -329,40 +329,48 @@ static int pnv_xive2_write_end(Xive2Router
>>> *xrtr, uint8_t blk, uint32_t idx,
>>> word_number);
>>> }
>>> -static int pnv_xive2_end_update(PnvXive2 *xive)
>>> +static int pnv_xive2_end_update(PnvXive2 *xive, uint8_t watch_engine)
>>> {
>>> - uint8_t blk = GETFIELD(VC_ENDC_WATCH_BLOCK_ID,
>>> - xive->vc_regs[(VC_ENDC_WATCH0_SPEC >> 3)]);
>>> - uint32_t idx = GETFIELD(VC_ENDC_WATCH_INDEX,
>>> - xive->vc_regs[(VC_ENDC_WATCH0_SPEC >> 3)]);
>>> - int i;
>>> + uint8_t blk;
>>> + uint32_t idx;
>>> + int i, spec_reg, data_reg;
>>> uint64_t endc_watch[4];
>>> + assert(watch_engine < ARRAY_SIZE(endc_watch));
>>> +
>>> + spec_reg = (VC_ENDC_WATCH0_SPEC + watch_engine * 0x40) >> 3;
>>> + data_reg = (VC_ENDC_WATCH0_DATA0 + watch_engine * 0x40) >> 3;
>>> + blk = GETFIELD(VC_ENDC_WATCH_BLOCK_ID, xive->vc_regs[spec_reg]);
>>> + idx = GETFIELD(VC_ENDC_WATCH_INDEX, xive->vc_regs[spec_reg]);
>>> +
>>> for (i = 0; i < ARRAY_SIZE(endc_watch); i++) {
>>> - endc_watch[i] =
>>> - cpu_to_be64(xive->vc_regs[(VC_ENDC_WATCH0_DATA0 >> 3) + i]);
>>> + endc_watch[i] = cpu_to_be64(xive->vc_regs[data_reg + i]);
>>> }
>>> return pnv_xive2_vst_write(xive, VST_END, blk, idx, endc_watch,
>>> XIVE_VST_WORD_ALL);
>>> }
>>> -static void pnv_xive2_end_cache_load(PnvXive2 *xive)
>>> +static void pnv_xive2_end_cache_load(PnvXive2 *xive, uint8_t
>>> watch_engine)
>>> {
>>> - uint8_t blk = GETFIELD(VC_ENDC_WATCH_BLOCK_ID,
>>> - xive->vc_regs[(VC_ENDC_WATCH0_SPEC >> 3)]);
>>> - uint32_t idx = GETFIELD(VC_ENDC_WATCH_INDEX,
>>> - xive->vc_regs[(VC_ENDC_WATCH0_SPEC >> 3)]);
>>> + uint8_t blk;
>>> + uint32_t idx;
>>> uint64_t endc_watch[4] = { 0 };
>>> - int i;
>>> + int i, spec_reg, data_reg;
>>> +
>>> + assert(watch_engine < ARRAY_SIZE(endc_watch));
>>> +
>>> + spec_reg = (VC_ENDC_WATCH0_SPEC + watch_engine * 0x40) >> 3;
>>> + data_reg = (VC_ENDC_WATCH0_DATA0 + watch_engine * 0x40) >> 3;
>>> + blk = GETFIELD(VC_ENDC_WATCH_BLOCK_ID, xive->vc_regs[spec_reg]);
>>> + idx = GETFIELD(VC_ENDC_WATCH_INDEX, xive->vc_regs[spec_reg]);
>>> if (pnv_xive2_vst_read(xive, VST_END, blk, idx, endc_watch)) {
>>> xive2_error(xive, "VST: no END entry %x/%x !?", blk, idx);
>>> }
>>> for (i = 0; i < ARRAY_SIZE(endc_watch); i++) {
>>> - xive->vc_regs[(VC_ENDC_WATCH0_DATA0 >> 3) + i] =
>>> - be64_to_cpu(endc_watch[i]);
>>> + xive->vc_regs[data_reg + i] = be64_to_cpu(endc_watch[i]);
>>> }
>>> }
>>> @@ -379,40 +387,50 @@ static int pnv_xive2_write_nvp(Xive2Router
>>> *xrtr, uint8_t blk, uint32_t idx,
>>> word_number);
>>> }
>>> -static int pnv_xive2_nvp_update(PnvXive2 *xive)
>>> +static int pnv_xive2_nvp_update(PnvXive2 *xive, uint8_t watch_engine)
>>> {
>>> - uint8_t blk = GETFIELD(PC_NXC_WATCH_BLOCK_ID,
>>> - xive->pc_regs[(PC_NXC_WATCH0_SPEC >> 3)]);
>>> - uint32_t idx = GETFIELD(PC_NXC_WATCH_INDEX,
>>> - xive->pc_regs[(PC_NXC_WATCH0_SPEC >> 3)]);
>>> - int i;
>>> + uint8_t blk;
>>> + uint32_t idx;
>>> + int i, spec_reg, data_reg;
>>> uint64_t nxc_watch[4];
>>> + if (watch_engine > 3) {
>>> + return -1;
>>> + }
>>
>> Why not an assert here and below ?
>>
>>
>> Thanks,
>>
>> C.
>>
> Hmmm, I added assert()s... Not sure what happened, maybe applying a
> different patch overwrote it. I will verify changes and add assert()s
> back in.
Strike that. The asserts were done in patch set 4.
>>
>>
>>
>>> + spec_reg = (PC_NXC_WATCH0_SPEC + watch_engine * 0x40) >> 3;
>>> + data_reg = (PC_NXC_WATCH0_DATA0 + watch_engine * 0x40) >> 3;
>>> + blk = GETFIELD(PC_NXC_WATCH_BLOCK_ID, xive->pc_regs[spec_reg]);
>>> + idx = GETFIELD(PC_NXC_WATCH_INDEX, xive->pc_regs[spec_reg]);
>>> +
>>> for (i = 0; i < ARRAY_SIZE(nxc_watch); i++) {
>>> - nxc_watch[i] =
>>> - cpu_to_be64(xive->pc_regs[(PC_NXC_WATCH0_DATA0 >> 3) +
>>> i]);
>>> + nxc_watch[i] = cpu_to_be64(xive->pc_regs[data_reg + i]);
>>> }
>>> return pnv_xive2_vst_write(xive, VST_NVP, blk, idx, nxc_watch,
>>> XIVE_VST_WORD_ALL);
>>> }
>>> -static void pnv_xive2_nvp_cache_load(PnvXive2 *xive)
>>> +static void pnv_xive2_nvp_cache_load(PnvXive2 *xive, uint8_t
>>> watch_engine)
>>> {
>>> - uint8_t blk = GETFIELD(PC_NXC_WATCH_BLOCK_ID,
>>> - xive->pc_regs[(PC_NXC_WATCH0_SPEC >> 3)]);
>>> - uint32_t idx = GETFIELD(PC_NXC_WATCH_INDEX,
>>> - xive->pc_regs[(PC_NXC_WATCH0_SPEC >> 3)]);
>>> + uint8_t blk;
>>> + uint32_t idx;
>>> uint64_t nxc_watch[4] = { 0 };
>>> - int i;
>>> + int i, spec_reg, data_reg;
>>> +
>>> + if (watch_engine > 3) {
>>> + return;
>>> + }
>>> + spec_reg = (PC_NXC_WATCH0_SPEC + watch_engine * 0x40) >> 3;
>>> + data_reg = (PC_NXC_WATCH0_DATA0 + watch_engine * 0x40) >> 3;
>>> + blk = GETFIELD(PC_NXC_WATCH_BLOCK_ID, xive->pc_regs[spec_reg]);
>>> + idx = GETFIELD(PC_NXC_WATCH_INDEX, xive->pc_regs[spec_reg]);
>>> if (pnv_xive2_vst_read(xive, VST_NVP, blk, idx, nxc_watch)) {
>>> xive2_error(xive, "VST: no NVP entry %x/%x !?", blk, idx);
>>> }
>>> for (i = 0; i < ARRAY_SIZE(nxc_watch); i++) {
>>> - xive->pc_regs[(PC_NXC_WATCH0_DATA0 >> 3) + i] =
>>> - be64_to_cpu(nxc_watch[i]);
>>> + xive->pc_regs[data_reg + i] = be64_to_cpu(nxc_watch[i]);
>>> }
>>> }
>>> @@ -964,12 +982,70 @@ static const MemoryRegionOps
>>> pnv_xive2_ic_cq_ops = {
>>> },
>>> };
>>> +static uint8_t pnv_xive2_cache_watch_assign(uint64_t engine_mask,
>>> + uint64_t *state)
>>> +{
>>> + uint8_t val = 0xFF;
>>> + int i;
>>> +
>>> + for (i = 3; i >= 0; i--) {
>>> + if (BIT(i) & engine_mask) {
>>> + if (!(BIT(i) & *state)) {
>>> + *state |= BIT(i);
>>> + val = 3 - i;
>>> + break;
>>> + }
>>> + }
>>> + }
>>> + return val;
>>> +}
>>> +
>>> +static void pnv_xive2_cache_watch_release(uint64_t *state, uint8_t
>>> watch_engine)
>>> +{
>>> + uint8_t engine_bit = 3 - watch_engine;
>>> +
>>> + if (*state & BIT(engine_bit)) {
>>> + *state &= ~BIT(engine_bit);
>>> + }
>>> +}
>>> +
>>> +static uint8_t pnv_xive2_endc_cache_watch_assign(PnvXive2 *xive)
>>> +{
>>> + uint64_t engine_mask = GETFIELD(VC_ENDC_CFG_CACHE_WATCH_ASSIGN,
>>> + xive->vc_regs[VC_ENDC_CFG >> 3]);
>>> + uint64_t state = xive->vc_regs[VC_ENDC_WATCH_ASSIGN >> 3];
>>> + uint8_t val;
>>> +
>>> + /*
>>> + * We keep track of which engines are currently busy in the
>>> + * VC_ENDC_WATCH_ASSIGN register directly. When the firmware reads
>>> + * the register, we don't return its value but the ID of an engine
>>> + * it can use.
>>> + * There are 4 engines. 0xFF means no engine is available.
>>> + */
>>> + val = pnv_xive2_cache_watch_assign(engine_mask, &state);
>>> + if (val != 0xFF) {
>>> + xive->vc_regs[VC_ENDC_WATCH_ASSIGN >> 3] = state;
>>> + }
>>> + return val;
>>> +}
>>> +
>>> +static void pnv_xive2_endc_cache_watch_release(PnvXive2 *xive,
>>> + uint8_t watch_engine)
>>> +{
>>> + uint64_t state = xive->vc_regs[VC_ENDC_WATCH_ASSIGN >> 3];
>>> +
>>> + pnv_xive2_cache_watch_release(&state, watch_engine);
>>> + xive->vc_regs[VC_ENDC_WATCH_ASSIGN >> 3] = state;
>>> +}
>>> +
>>> static uint64_t pnv_xive2_ic_vc_read(void *opaque, hwaddr offset,
>>> unsigned size)
>>> {
>>> PnvXive2 *xive = PNV_XIVE2(opaque);
>>> uint64_t val = 0;
>>> uint32_t reg = offset >> 3;
>>> + uint8_t watch_engine;
>>> switch (offset) {
>>> /*
>>> @@ -1000,24 +1076,44 @@ static uint64_t pnv_xive2_ic_vc_read(void
>>> *opaque, hwaddr offset,
>>> val = xive->vc_regs[reg];
>>> break;
>>> + case VC_ENDC_WATCH_ASSIGN:
>>> + val = pnv_xive2_endc_cache_watch_assign(xive);
>>> + break;
>>> +
>>> + case VC_ENDC_CFG:
>>> + val = xive->vc_regs[reg];
>>> + break;
>>> +
>>> /*
>>> * END cache updates
>>> */
>>> case VC_ENDC_WATCH0_SPEC:
>>> + case VC_ENDC_WATCH1_SPEC:
>>> + case VC_ENDC_WATCH2_SPEC:
>>> + case VC_ENDC_WATCH3_SPEC:
>>> + watch_engine = (offset - VC_ENDC_WATCH0_SPEC) >> 6;
>>> xive->vc_regs[reg] &= ~(VC_ENDC_WATCH_FULL |
>>> VC_ENDC_WATCH_CONFLICT);
>>> + pnv_xive2_endc_cache_watch_release(xive, watch_engine);
>>> val = xive->vc_regs[reg];
>>> break;
>>> case VC_ENDC_WATCH0_DATA0:
>>> + case VC_ENDC_WATCH1_DATA0:
>>> + case VC_ENDC_WATCH2_DATA0:
>>> + case VC_ENDC_WATCH3_DATA0:
>>> /*
>>> * Load DATA registers from cache with data requested by the
>>> * SPEC register
>>> */
>>> - pnv_xive2_end_cache_load(xive);
>>> + watch_engine = (offset - VC_ENDC_WATCH0_DATA0) >> 6;
>>> + pnv_xive2_end_cache_load(xive, watch_engine);
>>> val = xive->vc_regs[reg];
>>> break;
>>> case VC_ENDC_WATCH0_DATA1 ... VC_ENDC_WATCH0_DATA3:
>>> + case VC_ENDC_WATCH1_DATA1 ... VC_ENDC_WATCH1_DATA3:
>>> + case VC_ENDC_WATCH2_DATA1 ... VC_ENDC_WATCH2_DATA3:
>>> + case VC_ENDC_WATCH3_DATA1 ... VC_ENDC_WATCH3_DATA3:
>>> val = xive->vc_regs[reg];
>>> break;
>>> @@ -1063,6 +1159,7 @@ static void pnv_xive2_ic_vc_write(void
>>> *opaque, hwaddr offset,
>>> {
>>> PnvXive2 *xive = PNV_XIVE2(opaque);
>>> uint32_t reg = offset >> 3;
>>> + uint8_t watch_engine;
>>> switch (offset) {
>>> /*
>>> @@ -1095,19 +1192,32 @@ static void pnv_xive2_ic_vc_write(void
>>> *opaque, hwaddr offset,
>>> /* EAS update */
>>> break;
>>> + case VC_ENDC_CFG:
>>> + break;
>>> +
>>> /*
>>> * END cache updates
>>> */
>>> case VC_ENDC_WATCH0_SPEC:
>>> + case VC_ENDC_WATCH1_SPEC:
>>> + case VC_ENDC_WATCH2_SPEC:
>>> + case VC_ENDC_WATCH3_SPEC:
>>> val &= ~VC_ENDC_WATCH_CONFLICT; /* HW will set this bit */
>>> break;
>>> case VC_ENDC_WATCH0_DATA1 ... VC_ENDC_WATCH0_DATA3:
>>> + case VC_ENDC_WATCH1_DATA1 ... VC_ENDC_WATCH1_DATA3:
>>> + case VC_ENDC_WATCH2_DATA1 ... VC_ENDC_WATCH2_DATA3:
>>> + case VC_ENDC_WATCH3_DATA1 ... VC_ENDC_WATCH3_DATA3:
>>> break;
>>> case VC_ENDC_WATCH0_DATA0:
>>> + case VC_ENDC_WATCH1_DATA0:
>>> + case VC_ENDC_WATCH2_DATA0:
>>> + case VC_ENDC_WATCH3_DATA0:
>>> /* writing to DATA0 triggers the cache write */
>>> + watch_engine = (offset - VC_ENDC_WATCH0_DATA0) >> 6;
>>> xive->vc_regs[reg] = val;
>>> - pnv_xive2_end_update(xive);
>>> + pnv_xive2_end_update(xive, watch_engine);
>>> break;
>>> @@ -1157,12 +1267,43 @@ static const MemoryRegionOps
>>> pnv_xive2_ic_vc_ops = {
>>> },
>>> };
>>> +static uint8_t pnv_xive2_nxc_cache_watch_assign(PnvXive2 *xive)
>>> +{
>>> + uint64_t engine_mask = GETFIELD(PC_NXC_PROC_CONFIG_WATCH_ASSIGN,
>>> + xive->pc_regs[PC_NXC_PROC_CONFIG >> 3]);
>>> + uint64_t state = xive->pc_regs[PC_NXC_WATCH_ASSIGN >> 3];
>>> + uint8_t val;
>>> +
>>> + /*
>>> + * We keep track of which engines are currently busy in the
>>> + * PC_NXC_WATCH_ASSIGN register directly. When the firmware reads
>>> + * the register, we don't return its value but the ID of an engine
>>> + * it can use.
>>> + * There are 4 engines. 0xFF means no engine is available.
>>> + */
>>> + val = pnv_xive2_cache_watch_assign(engine_mask, &state);
>>> + if (val != 0xFF) {
>>> + xive->pc_regs[PC_NXC_WATCH_ASSIGN >> 3] = state;
>>> + }
>>> + return val;
>>> +}
>>> +
>>> +static void pnv_xive2_nxc_cache_watch_release(PnvXive2 *xive,
>>> + uint8_t watch_engine)
>>> +{
>>> + uint64_t state = xive->pc_regs[PC_NXC_WATCH_ASSIGN >> 3];
>>> +
>>> + pnv_xive2_cache_watch_release(&state, watch_engine);
>>> + xive->pc_regs[PC_NXC_WATCH_ASSIGN >> 3] = state;
>>> +}
>>> +
>>> static uint64_t pnv_xive2_ic_pc_read(void *opaque, hwaddr offset,
>>> unsigned size)
>>> {
>>> PnvXive2 *xive = PNV_XIVE2(opaque);
>>> uint64_t val = -1;
>>> uint32_t reg = offset >> 3;
>>> + uint8_t watch_engine;
>>> switch (offset) {
>>> /*
>>> @@ -1173,24 +1314,44 @@ static uint64_t pnv_xive2_ic_pc_read(void
>>> *opaque, hwaddr offset,
>>> val = xive->pc_regs[reg];
>>> break;
>>> + case PC_NXC_WATCH_ASSIGN:
>>> + val = pnv_xive2_nxc_cache_watch_assign(xive);
>>> + break;
>>> +
>>> + case PC_NXC_PROC_CONFIG:
>>> + val = xive->pc_regs[reg];
>>> + break;
>>> +
>>> /*
>>> * cache updates
>>> */
>>> case PC_NXC_WATCH0_SPEC:
>>> + case PC_NXC_WATCH1_SPEC:
>>> + case PC_NXC_WATCH2_SPEC:
>>> + case PC_NXC_WATCH3_SPEC:
>>> + watch_engine = (offset - PC_NXC_WATCH0_SPEC) >> 6;
>>> xive->pc_regs[reg] &= ~(PC_NXC_WATCH_FULL |
>>> PC_NXC_WATCH_CONFLICT);
>>> + pnv_xive2_nxc_cache_watch_release(xive, watch_engine);
>>> val = xive->pc_regs[reg];
>>> break;
>>> case PC_NXC_WATCH0_DATA0:
>>> + case PC_NXC_WATCH1_DATA0:
>>> + case PC_NXC_WATCH2_DATA0:
>>> + case PC_NXC_WATCH3_DATA0:
>>> /*
>>> * Load DATA registers from cache with data requested by the
>>> * SPEC register
>>> */
>>> - pnv_xive2_nvp_cache_load(xive);
>>> + watch_engine = (offset - PC_NXC_WATCH0_DATA0) >> 6;
>>> + pnv_xive2_nvp_cache_load(xive, watch_engine);
>>> val = xive->pc_regs[reg];
>>> break;
>>> case PC_NXC_WATCH0_DATA1 ... PC_NXC_WATCH0_DATA3:
>>> + case PC_NXC_WATCH1_DATA1 ... PC_NXC_WATCH1_DATA3:
>>> + case PC_NXC_WATCH2_DATA1 ... PC_NXC_WATCH2_DATA3:
>>> + case PC_NXC_WATCH3_DATA1 ... PC_NXC_WATCH3_DATA3:
>>> val = xive->pc_regs[reg];
>>> break;
>>> @@ -1219,6 +1380,7 @@ static void pnv_xive2_ic_pc_write(void
>>> *opaque, hwaddr offset,
>>> {
>>> PnvXive2 *xive = PNV_XIVE2(opaque);
>>> uint32_t reg = offset >> 3;
>>> + uint8_t watch_engine;
>>> switch (offset) {
>>> @@ -1231,19 +1393,32 @@ static void pnv_xive2_ic_pc_write(void
>>> *opaque, hwaddr offset,
>>> case PC_VSD_TABLE_DATA:
>>> break;
>>> + case PC_NXC_PROC_CONFIG:
>>> + break;
>>> +
>>> /*
>>> * cache updates
>>> */
>>> case PC_NXC_WATCH0_SPEC:
>>> + case PC_NXC_WATCH1_SPEC:
>>> + case PC_NXC_WATCH2_SPEC:
>>> + case PC_NXC_WATCH3_SPEC:
>>> val &= ~PC_NXC_WATCH_CONFLICT; /* HW will set this bit */
>>> break;
>>> case PC_NXC_WATCH0_DATA1 ... PC_NXC_WATCH0_DATA3:
>>> + case PC_NXC_WATCH1_DATA1 ... PC_NXC_WATCH1_DATA3:
>>> + case PC_NXC_WATCH2_DATA1 ... PC_NXC_WATCH2_DATA3:
>>> + case PC_NXC_WATCH3_DATA1 ... PC_NXC_WATCH3_DATA3:
>>> break;
>>> case PC_NXC_WATCH0_DATA0:
>>> + case PC_NXC_WATCH1_DATA0:
>>> + case PC_NXC_WATCH2_DATA0:
>>> + case PC_NXC_WATCH3_DATA0:
>>> /* writing to DATA0 triggers the cache write */
>>> + watch_engine = (offset - PC_NXC_WATCH0_DATA0) >> 6;
>>> xive->pc_regs[reg] = val;
>>> - pnv_xive2_nvp_update(xive);
>>> + pnv_xive2_nvp_update(xive, watch_engine);
>>> break;
>>> /* case PC_NXC_FLUSH_CTRL: */
>>> @@ -1814,6 +1989,12 @@ static void pnv_xive2_reset(void *dev)
>>> xive->cq_regs[CQ_XIVE_CFG >> 3] |=
>>> SETFIELD(CQ_XIVE_CFG_HYP_HARD_BLOCK_ID, 0ull,
>>> xive->chip->chip_id);
>>> + /* VC and PC cache watch assign mechanism */
>>> + xive->vc_regs[VC_ENDC_CFG >> 3] =
>>> + SETFIELD(VC_ENDC_CFG_CACHE_WATCH_ASSIGN, 0ull, 0b0111);
>>> + xive->pc_regs[PC_NXC_PROC_CONFIG >> 3] =
>>> + SETFIELD(PC_NXC_PROC_CONFIG_WATCH_ASSIGN, 0ull, 0b0111);
>>> +
>>> /* Set default page size to 64k */
>>> xive->ic_shift = xive->esb_shift = xive->end_shift = 16;
>>> xive->nvc_shift = xive->nvpg_shift = xive->tm_shift = 16;
>>
^ permalink raw reply [flat|nested] 23+ messages in thread
* [PATCH v2 2/9] pnv/xive2: Structure/define alignment changes
2024-07-15 18:33 [PATCH v2 0/9] XIVE changes for Cache Watch, VSTs, STT and info pic Michael Kowal
2024-07-15 18:33 ` [PATCH v2 1/9] pnv/xive2: XIVE2 Cache Watch, Cache Flush and Sync Injection support Michael Kowal
@ 2024-07-15 18:33 ` Michael Kowal
2024-07-16 7:04 ` Cédric Le Goater
2024-07-15 18:33 ` [PATCH v2 3/9] pnv/xive: Support cache flush and queue sync inject with notifications Michael Kowal
` (6 subsequent siblings)
8 siblings, 1 reply; 23+ messages in thread
From: Michael Kowal @ 2024-07-15 18:33 UTC (permalink / raw)
To: qemu-devel; +Cc: qemu-ppc, clg, fbarrat, npiggin, milesg
Made changes to some structure and define elements to ease review in
next patchset.
Signed-off-by: Michael Kowal <kowal@linux.vnet.ibm.com>
---
hw/intc/pnv_xive2.c | 30 +++++++++++++++---------------
1 file changed, 15 insertions(+), 15 deletions(-)
diff --git a/hw/intc/pnv_xive2.c b/hw/intc/pnv_xive2.c
index f6a735cca5..5b075a220c 100644
--- a/hw/intc/pnv_xive2.c
+++ b/hw/intc/pnv_xive2.c
@@ -45,16 +45,16 @@ typedef struct XiveVstInfo {
static const XiveVstInfo vst_infos[] = {
- [VST_EAS] = { "EAT", sizeof(Xive2Eas), 16 },
- [VST_ESB] = { "ESB", 1, 16 },
- [VST_END] = { "ENDT", sizeof(Xive2End), 16 },
+ [VST_EAS] = { "EAT", sizeof(Xive2Eas), 16 },
+ [VST_ESB] = { "ESB", 1, 16 },
+ [VST_END] = { "ENDT", sizeof(Xive2End), 16 },
- [VST_NVP] = { "NVPT", sizeof(Xive2Nvp), 16 },
- [VST_NVG] = { "NVGT", sizeof(Xive2Nvgc), 16 },
- [VST_NVC] = { "NVCT", sizeof(Xive2Nvgc), 16 },
+ [VST_NVP] = { "NVPT", sizeof(Xive2Nvp), 16 },
+ [VST_NVG] = { "NVGT", sizeof(Xive2Nvgc), 16 },
+ [VST_NVC] = { "NVCT", sizeof(Xive2Nvgc), 16 },
- [VST_IC] = { "IC", 1 /* ? */ , 16 }, /* Topology # */
- [VST_SYNC] = { "SYNC", 1 /* ? */ , 16 }, /* Topology # */
+ [VST_IC] = { "IC", 1, /* ? */ 16 }, /* Topology # */
+ [VST_SYNC] = { "SYNC", 1, /* ? */ 16 }, /* Topology # */
/*
* This table contains the backing store pages for the interrupt
@@ -1722,13 +1722,13 @@ static const MemoryRegionOps pnv_xive2_ic_lsi_ops = {
/*
* Sync MMIO page (write only)
*/
-#define PNV_XIVE2_SYNC_IPI 0x000
-#define PNV_XIVE2_SYNC_HW 0x080
-#define PNV_XIVE2_SYNC_NxC 0x100
-#define PNV_XIVE2_SYNC_INT 0x180
-#define PNV_XIVE2_SYNC_OS_ESC 0x200
-#define PNV_XIVE2_SYNC_POOL_ESC 0x280
-#define PNV_XIVE2_SYNC_HARD_ESC 0x300
+#define PNV_XIVE2_SYNC_IPI 0x000
+#define PNV_XIVE2_SYNC_HW 0x080
+#define PNV_XIVE2_SYNC_NxC 0x100
+#define PNV_XIVE2_SYNC_INT 0x180
+#define PNV_XIVE2_SYNC_OS_ESC 0x200
+#define PNV_XIVE2_SYNC_POOL_ESC 0x280
+#define PNV_XIVE2_SYNC_HARD_ESC 0x300
static uint64_t pnv_xive2_ic_sync_read(void *opaque, hwaddr offset,
unsigned size)
--
2.43.0
^ permalink raw reply related [flat|nested] 23+ messages in thread* Re: [PATCH v2 2/9] pnv/xive2: Structure/define alignment changes
2024-07-15 18:33 ` [PATCH v2 2/9] pnv/xive2: Structure/define alignment changes Michael Kowal
@ 2024-07-16 7:04 ` Cédric Le Goater
0 siblings, 0 replies; 23+ messages in thread
From: Cédric Le Goater @ 2024-07-16 7:04 UTC (permalink / raw)
To: Michael Kowal, qemu-devel; +Cc: qemu-ppc, fbarrat, npiggin, milesg
On 7/15/24 20:33, Michael Kowal wrote:
> Made changes to some structure and define elements to ease review in
> next patchset.
>
> Signed-off-by: Michael Kowal <kowal@linux.vnet.ibm.com>
Reviewed-by: Cédric Le Goater <clg@redhat.com>
Thanks,
C.
> ---
> hw/intc/pnv_xive2.c | 30 +++++++++++++++---------------
> 1 file changed, 15 insertions(+), 15 deletions(-)
>
> diff --git a/hw/intc/pnv_xive2.c b/hw/intc/pnv_xive2.c
> index f6a735cca5..5b075a220c 100644
> --- a/hw/intc/pnv_xive2.c
> +++ b/hw/intc/pnv_xive2.c
> @@ -45,16 +45,16 @@ typedef struct XiveVstInfo {
>
> static const XiveVstInfo vst_infos[] = {
>
> - [VST_EAS] = { "EAT", sizeof(Xive2Eas), 16 },
> - [VST_ESB] = { "ESB", 1, 16 },
> - [VST_END] = { "ENDT", sizeof(Xive2End), 16 },
> + [VST_EAS] = { "EAT", sizeof(Xive2Eas), 16 },
> + [VST_ESB] = { "ESB", 1, 16 },
> + [VST_END] = { "ENDT", sizeof(Xive2End), 16 },
>
> - [VST_NVP] = { "NVPT", sizeof(Xive2Nvp), 16 },
> - [VST_NVG] = { "NVGT", sizeof(Xive2Nvgc), 16 },
> - [VST_NVC] = { "NVCT", sizeof(Xive2Nvgc), 16 },
> + [VST_NVP] = { "NVPT", sizeof(Xive2Nvp), 16 },
> + [VST_NVG] = { "NVGT", sizeof(Xive2Nvgc), 16 },
> + [VST_NVC] = { "NVCT", sizeof(Xive2Nvgc), 16 },
>
> - [VST_IC] = { "IC", 1 /* ? */ , 16 }, /* Topology # */
> - [VST_SYNC] = { "SYNC", 1 /* ? */ , 16 }, /* Topology # */
> + [VST_IC] = { "IC", 1, /* ? */ 16 }, /* Topology # */
> + [VST_SYNC] = { "SYNC", 1, /* ? */ 16 }, /* Topology # */
>
> /*
> * This table contains the backing store pages for the interrupt
> @@ -1722,13 +1722,13 @@ static const MemoryRegionOps pnv_xive2_ic_lsi_ops = {
> /*
> * Sync MMIO page (write only)
> */
> -#define PNV_XIVE2_SYNC_IPI 0x000
> -#define PNV_XIVE2_SYNC_HW 0x080
> -#define PNV_XIVE2_SYNC_NxC 0x100
> -#define PNV_XIVE2_SYNC_INT 0x180
> -#define PNV_XIVE2_SYNC_OS_ESC 0x200
> -#define PNV_XIVE2_SYNC_POOL_ESC 0x280
> -#define PNV_XIVE2_SYNC_HARD_ESC 0x300
> +#define PNV_XIVE2_SYNC_IPI 0x000
> +#define PNV_XIVE2_SYNC_HW 0x080
> +#define PNV_XIVE2_SYNC_NxC 0x100
> +#define PNV_XIVE2_SYNC_INT 0x180
> +#define PNV_XIVE2_SYNC_OS_ESC 0x200
> +#define PNV_XIVE2_SYNC_POOL_ESC 0x280
> +#define PNV_XIVE2_SYNC_HARD_ESC 0x300
>
> static uint64_t pnv_xive2_ic_sync_read(void *opaque, hwaddr offset,
> unsigned size)
^ permalink raw reply [flat|nested] 23+ messages in thread
* [PATCH v2 3/9] pnv/xive: Support cache flush and queue sync inject with notifications
2024-07-15 18:33 [PATCH v2 0/9] XIVE changes for Cache Watch, VSTs, STT and info pic Michael Kowal
2024-07-15 18:33 ` [PATCH v2 1/9] pnv/xive2: XIVE2 Cache Watch, Cache Flush and Sync Injection support Michael Kowal
2024-07-15 18:33 ` [PATCH v2 2/9] pnv/xive2: Structure/define alignment changes Michael Kowal
@ 2024-07-15 18:33 ` Michael Kowal
2024-07-16 7:26 ` Cédric Le Goater
2024-07-15 18:33 ` [PATCH v2 4/9] pnv/xive2: Add NVG and NVC to cache watch facility Michael Kowal
` (5 subsequent siblings)
8 siblings, 1 reply; 23+ messages in thread
From: Michael Kowal @ 2024-07-15 18:33 UTC (permalink / raw)
To: qemu-devel; +Cc: qemu-ppc, clg, fbarrat, npiggin, milesg
From: Nicholas Piggin <npiggin@gmail,com>
Adds support for writing a completion notification byte in memory
whenever a cache flush or queue sync inject operation is requested by
software. QEMU does not cache any of the XIVE data that is in memory and
therefore it simply writes the completion notification byte at the time
that the operation is requested.
Co-authored-by: Glenn Miles <milesg@linux.vnet.ibm.com>
Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Glenn Miles <milesg@linux.vnet.ibm.com>
Signed-off-by: Michael Kowal <kowal@linux.vnet.ibm.com>
---
hw/intc/pnv_xive2_regs.h | 16 ++++
include/hw/ppc/pnv_chip.h | 1 +
hw/intc/pnv_xive2.c | 151 +++++++++++++++++++++++++++++++++++++-
3 files changed, 166 insertions(+), 2 deletions(-)
diff --git a/hw/intc/pnv_xive2_regs.h b/hw/intc/pnv_xive2_regs.h
index f8e4a677c6..ca05255d20 100644
--- a/hw/intc/pnv_xive2_regs.h
+++ b/hw/intc/pnv_xive2_regs.h
@@ -232,6 +232,10 @@
#define VC_ESBC_FLUSH_POLL_BLOCK_ID_MASK PPC_BITMASK(32, 35)
#define VC_ESBC_FLUSH_POLL_OFFSET_MASK PPC_BITMASK(36, 63) /* 28-bit */
+/* ESBC cache flush inject register */
+#define X_VC_ESBC_FLUSH_INJECT 0x142
+#define VC_ESBC_FLUSH_INJECT 0x210
+
/* ESBC configuration */
#define X_VC_ESBC_CFG 0x148
#define VC_ESBC_CFG 0x240
@@ -250,6 +254,10 @@
#define VC_EASC_FLUSH_POLL_BLOCK_ID_MASK PPC_BITMASK(32, 35)
#define VC_EASC_FLUSH_POLL_OFFSET_MASK PPC_BITMASK(36, 63) /* 28-bit */
+/* EASC flush inject register */
+#define X_VC_EASC_FLUSH_INJECT 0x162
+#define VC_EASC_FLUSH_INJECT 0x310
+
/*
* VC2
*/
@@ -270,6 +278,10 @@
#define VC_ENDC_FLUSH_POLL_BLOCK_ID_MASK PPC_BITMASK(36, 39)
#define VC_ENDC_FLUSH_POLL_OFFSET_MASK PPC_BITMASK(40, 63) /* 24-bit */
+/* ENDC flush inject register */
+#define X_VC_ENDC_FLUSH_INJECT 0x182
+#define VC_ENDC_FLUSH_INJECT 0x410
+
/* ENDC Sync done */
#define X_VC_ENDC_SYNC_DONE 0x184
#define VC_ENDC_SYNC_DONE 0x420
@@ -403,6 +415,10 @@
#define PC_NXC_FLUSH_POLL_BLOCK_ID_MASK PPC_BITMASK(36, 39)
#define PC_NXC_FLUSH_POLL_OFFSET_MASK PPC_BITMASK(40, 63) /* 24-bit */
+/* NxC Cache flush inject */
+#define X_PC_NXC_FLUSH_INJECT 0x282
+#define PC_NXC_FLUSH_INJECT 0x410
+
/* NxC Cache watch assign */
#define X_PC_NXC_WATCH_ASSIGN 0x286
#define PC_NXC_WATCH_ASSIGN 0x430
diff --git a/include/hw/ppc/pnv_chip.h b/include/hw/ppc/pnv_chip.h
index a4ed17ac59..9502248561 100644
--- a/include/hw/ppc/pnv_chip.h
+++ b/include/hw/ppc/pnv_chip.h
@@ -131,6 +131,7 @@ struct Pnv10Chip {
#define PNV10_PIR2FUSEDCORE(pir) (((pir) >> 3) & 0xf)
#define PNV10_PIR2CHIP(pir) (((pir) >> 8) & 0x7f)
+#define PNV10_PIR2THREAD(pir) (((pir) & 0x7f))
struct PnvChipClass {
/*< private >*/
diff --git a/hw/intc/pnv_xive2.c b/hw/intc/pnv_xive2.c
index 5b075a220c..d62ac20d98 100644
--- a/hw/intc/pnv_xive2.c
+++ b/hw/intc/pnv_xive2.c
@@ -25,6 +25,7 @@
#include "hw/ppc/ppc.h"
#include "hw/qdev-properties.h"
#include "sysemu/reset.h"
+#include "sysemu/qtest.h"
#include <libfdt.h>
@@ -32,6 +33,16 @@
#undef XIVE2_DEBUG
+/* XIVE Sync or Flush Notification Block */
+typedef struct XiveSfnBlock {
+ uint8_t bytes[32];
+} XiveSfnBlock;
+
+/* XIVE Thread Sync or Flush Notification Area */
+typedef struct XiveThreadNA {
+ XiveSfnBlock topo[16];
+} XiveThreadNA;
+
/*
* Virtual structures table (VST)
*/
@@ -54,7 +65,7 @@ static const XiveVstInfo vst_infos[] = {
[VST_NVC] = { "NVCT", sizeof(Xive2Nvgc), 16 },
[VST_IC] = { "IC", 1, /* ? */ 16 }, /* Topology # */
- [VST_SYNC] = { "SYNC", 1, /* ? */ 16 }, /* Topology # */
+ [VST_SYNC] = { "SYNC", sizeof(XiveThreadNA), 16 }, /* Topology # */
/*
* This table contains the backing store pages for the interrupt
@@ -329,6 +340,70 @@ static int pnv_xive2_write_end(Xive2Router *xrtr, uint8_t blk, uint32_t idx,
word_number);
}
+static inline int pnv_xive2_get_current_pir(PnvXive2 *xive)
+{
+ if (!qtest_enabled()) {
+ PowerPCCPU *cpu = POWERPC_CPU(current_cpu);
+ return ppc_cpu_pir(cpu);
+ }
+ return 0;
+}
+
+/*
+ * After SW injects a Queue Sync or Cache Flush operation, HW will notify
+ * SW of the completion of the operation by writing a byte of all 1's (0xff)
+ * to a specific memory location. The memory location is calculated by first
+ * looking up a base address in the SYNC VSD using the Topology ID of the
+ * originating thread as the "block" number. This points to a
+ * 64k block of memory that is further divided into 128 512 byte chunks of
+ * memory, which is indexed by the thread id of the requesting thread.
+ * Finally, this 512 byte chunk of memory is divided into 16 32 byte
+ * chunks which are indexed by the topology id of the targeted IC's chip.
+ * The values below are the offsets into that 32 byte chunk of memory for
+ * each type of cache flush or queue sync operation.
+ */
+#define PNV_XIVE2_QUEUE_IPI 0x00
+#define PNV_XIVE2_QUEUE_HW 0x01
+#define PNV_XIVE2_QUEUE_NXC 0x02
+#define PNV_XIVE2_QUEUE_INT 0x03
+#define PNV_XIVE2_QUEUE_OS 0x04
+#define PNV_XIVE2_QUEUE_POOL 0x05
+#define PNV_XIVE2_QUEUE_HARD 0x06
+#define PNV_XIVE2_CACHE_ENDC 0x08
+#define PNV_XIVE2_CACHE_ESBC 0x09
+#define PNV_XIVE2_CACHE_EASC 0x0a
+#define PNV_XIVE2_QUEUE_NXC_LD_LCL_NCO 0x10
+#define PNV_XIVE2_QUEUE_NXC_LD_LCL_CO 0x11
+#define PNV_XIVE2_QUEUE_NXC_ST_LCL_NCI 0x12
+#define PNV_XIVE2_QUEUE_NXC_ST_LCL_CI 0x13
+#define PNV_XIVE2_QUEUE_NXC_ST_RMT_NCI 0x14
+#define PNV_XIVE2_QUEUE_NXC_ST_RMT_CI 0x15
+#define PNV_XIVE2_CACHE_NXC 0x18
+
+static int pnv_xive2_inject_notify(PnvXive2 *xive, int type)
+{
+ uint64_t addr;
+ int pir = pnv_xive2_get_current_pir(xive);
+ int thread_nr = PNV10_PIR2THREAD(pir);
+ int thread_topo_id = PNV10_PIR2CHIP(pir);
+ int ic_topo_id = xive->chip->chip_id;
+ uint64_t offset = ic_topo_id * sizeof(XiveSfnBlock);
+ uint8_t byte = 0xff;
+
+ /* Retrieve the address of requesting thread's notification area */
+ addr = pnv_xive2_vst_addr(xive, VST_SYNC, thread_topo_id, thread_nr);
+
+ if (!addr) {
+ xive2_error(xive, "VST: no SYNC entry %x/%x !?",
+ thread_topo_id, thread_nr);
+ return -1;
+ }
+
+ cpu_physical_memory_write(addr + offset + type, &byte, 1);
+
+ return 0;
+}
+
static int pnv_xive2_end_update(PnvXive2 *xive, uint8_t watch_engine)
{
uint8_t blk;
@@ -1180,6 +1255,10 @@ static void pnv_xive2_ic_vc_write(void *opaque, hwaddr offset,
/* ESB update */
break;
+ case VC_ESBC_FLUSH_INJECT:
+ pnv_xive2_inject_notify(xive, PNV_XIVE2_CACHE_ESBC);
+ break;
+
case VC_ESBC_CFG:
break;
@@ -1192,6 +1271,10 @@ static void pnv_xive2_ic_vc_write(void *opaque, hwaddr offset,
/* EAS update */
break;
+ case VC_EASC_FLUSH_INJECT:
+ pnv_xive2_inject_notify(xive, PNV_XIVE2_CACHE_EASC);
+ break;
+
case VC_ENDC_CFG:
break;
@@ -1226,6 +1309,10 @@ static void pnv_xive2_ic_vc_write(void *opaque, hwaddr offset,
xive->vc_regs[VC_ENDC_FLUSH_CTRL >> 3] |= VC_ENDC_FLUSH_CTRL_POLL_VALID;
break;
+ case VC_ENDC_FLUSH_INJECT:
+ pnv_xive2_inject_notify(xive, PNV_XIVE2_CACHE_ENDC);
+ break;
+
/*
* Indirect invalidation
*/
@@ -1426,6 +1513,10 @@ static void pnv_xive2_ic_pc_write(void *opaque, hwaddr offset,
xive->pc_regs[PC_NXC_FLUSH_CTRL >> 3] |= PC_NXC_FLUSH_CTRL_POLL_VALID;
break;
+ case PC_NXC_FLUSH_INJECT:
+ pnv_xive2_inject_notify(xive, PNV_XIVE2_CACHE_NXC);
+ break;
+
/*
* Indirect invalidation
*/
@@ -1729,6 +1820,12 @@ static const MemoryRegionOps pnv_xive2_ic_lsi_ops = {
#define PNV_XIVE2_SYNC_OS_ESC 0x200
#define PNV_XIVE2_SYNC_POOL_ESC 0x280
#define PNV_XIVE2_SYNC_HARD_ESC 0x300
+#define PNV_XIVE2_SYNC_NXC_LD_LCL_NCO 0x800
+#define PNV_XIVE2_SYNC_NXC_LD_LCL_CO 0x880
+#define PNV_XIVE2_SYNC_NXC_ST_LCL_NCI 0x900
+#define PNV_XIVE2_SYNC_NXC_ST_LCL_CI 0x980
+#define PNV_XIVE2_SYNC_NXC_ST_RMT_NCI 0xA00
+#define PNV_XIVE2_SYNC_NXC_ST_RMT_CI 0xA80
static uint64_t pnv_xive2_ic_sync_read(void *opaque, hwaddr offset,
unsigned size)
@@ -1740,22 +1837,72 @@ static uint64_t pnv_xive2_ic_sync_read(void *opaque, hwaddr offset,
return -1;
}
+/*
+ * The sync MMIO space spans two pages. The lower page is use for
+ * queue sync "poll" requests while the upper page is used for queue
+ * sync "inject" requests. Inject requests require the HW to write
+ * a byte of all 1's to a predetermined location in memory in order
+ * to signal completion of the request. Both pages have the same
+ * layout, so it is easiest to handle both with a single function.
+ */
static void pnv_xive2_ic_sync_write(void *opaque, hwaddr offset,
uint64_t val, unsigned size)
{
PnvXive2 *xive = PNV_XIVE2(opaque);
+ int inject_type;
+ hwaddr pg_offset_mask = (1ull << xive->ic_shift) - 1;
- switch (offset) {
+ /* adjust offset for inject page */
+ hwaddr adj_offset = offset & pg_offset_mask;
+
+ switch (adj_offset) {
case PNV_XIVE2_SYNC_IPI:
+ inject_type = PNV_XIVE2_QUEUE_IPI;
+ break;
case PNV_XIVE2_SYNC_HW:
+ inject_type = PNV_XIVE2_QUEUE_HW;
+ break;
case PNV_XIVE2_SYNC_NxC:
+ inject_type = PNV_XIVE2_QUEUE_NXC;
+ break;
case PNV_XIVE2_SYNC_INT:
+ inject_type = PNV_XIVE2_QUEUE_INT;
+ break;
case PNV_XIVE2_SYNC_OS_ESC:
+ inject_type = PNV_XIVE2_QUEUE_OS;
+ break;
case PNV_XIVE2_SYNC_POOL_ESC:
+ inject_type = PNV_XIVE2_QUEUE_POOL;
+ break;
case PNV_XIVE2_SYNC_HARD_ESC:
+ inject_type = PNV_XIVE2_QUEUE_HARD;
+ break;
+ case PNV_XIVE2_SYNC_NXC_LD_LCL_NCO:
+ inject_type = PNV_XIVE2_QUEUE_NXC_LD_LCL_NCO;
+ break;
+ case PNV_XIVE2_SYNC_NXC_LD_LCL_CO:
+ inject_type = PNV_XIVE2_QUEUE_NXC_LD_LCL_CO;
+ break;
+ case PNV_XIVE2_SYNC_NXC_ST_LCL_NCI:
+ inject_type = PNV_XIVE2_QUEUE_NXC_ST_LCL_NCI;
+ break;
+ case PNV_XIVE2_SYNC_NXC_ST_LCL_CI:
+ inject_type = PNV_XIVE2_QUEUE_NXC_ST_LCL_CI;
+ break;
+ case PNV_XIVE2_SYNC_NXC_ST_RMT_NCI:
+ inject_type = PNV_XIVE2_QUEUE_NXC_ST_RMT_NCI;
+ break;
+ case PNV_XIVE2_SYNC_NXC_ST_RMT_CI:
+ inject_type = PNV_XIVE2_QUEUE_NXC_ST_RMT_CI;
break;
default:
xive2_error(xive, "SYNC: invalid write @%"HWADDR_PRIx, offset);
+ return;
+ }
+
+ /* Write Queue Sync notification byte if writing to sync inject page */
+ if ((offset & ~pg_offset_mask) != 0) {
+ pnv_xive2_inject_notify(xive, inject_type);
}
}
--
2.43.0
^ permalink raw reply related [flat|nested] 23+ messages in thread* Re: [PATCH v2 3/9] pnv/xive: Support cache flush and queue sync inject with notifications
2024-07-15 18:33 ` [PATCH v2 3/9] pnv/xive: Support cache flush and queue sync inject with notifications Michael Kowal
@ 2024-07-16 7:26 ` Cédric Le Goater
0 siblings, 0 replies; 23+ messages in thread
From: Cédric Le Goater @ 2024-07-16 7:26 UTC (permalink / raw)
To: Michael Kowal, qemu-devel; +Cc: qemu-ppc, fbarrat, npiggin, milesg
On 7/15/24 20:33, Michael Kowal wrote:
> From: Nicholas Piggin <npiggin@gmail,com>
>
> Adds support for writing a completion notification byte in memory
> whenever a cache flush or queue sync inject operation is requested by
> software. QEMU does not cache any of the XIVE data that is in memory and
> therefore it simply writes the completion notification byte at the time
> that the operation is requested.
>
> Co-authored-by: Glenn Miles <milesg@linux.vnet.ibm.com>
>
> Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
> Signed-off-by: Glenn Miles <milesg@linux.vnet.ibm.com>
> Signed-off-by: Michael Kowal <kowal@linux.vnet.ibm.com>
> ---
> hw/intc/pnv_xive2_regs.h | 16 ++++
> include/hw/ppc/pnv_chip.h | 1 +
> hw/intc/pnv_xive2.c | 151 +++++++++++++++++++++++++++++++++++++-
> 3 files changed, 166 insertions(+), 2 deletions(-)
>
> diff --git a/hw/intc/pnv_xive2_regs.h b/hw/intc/pnv_xive2_regs.h
> index f8e4a677c6..ca05255d20 100644
> --- a/hw/intc/pnv_xive2_regs.h
> +++ b/hw/intc/pnv_xive2_regs.h
> @@ -232,6 +232,10 @@
> #define VC_ESBC_FLUSH_POLL_BLOCK_ID_MASK PPC_BITMASK(32, 35)
> #define VC_ESBC_FLUSH_POLL_OFFSET_MASK PPC_BITMASK(36, 63) /* 28-bit */
>
> +/* ESBC cache flush inject register */
> +#define X_VC_ESBC_FLUSH_INJECT 0x142
> +#define VC_ESBC_FLUSH_INJECT 0x210
> +
> /* ESBC configuration */
> #define X_VC_ESBC_CFG 0x148
> #define VC_ESBC_CFG 0x240
> @@ -250,6 +254,10 @@
> #define VC_EASC_FLUSH_POLL_BLOCK_ID_MASK PPC_BITMASK(32, 35)
> #define VC_EASC_FLUSH_POLL_OFFSET_MASK PPC_BITMASK(36, 63) /* 28-bit */
>
> +/* EASC flush inject register */
> +#define X_VC_EASC_FLUSH_INJECT 0x162
> +#define VC_EASC_FLUSH_INJECT 0x310
> +
> /*
> * VC2
> */
> @@ -270,6 +278,10 @@
> #define VC_ENDC_FLUSH_POLL_BLOCK_ID_MASK PPC_BITMASK(36, 39)
> #define VC_ENDC_FLUSH_POLL_OFFSET_MASK PPC_BITMASK(40, 63) /* 24-bit */
>
> +/* ENDC flush inject register */
> +#define X_VC_ENDC_FLUSH_INJECT 0x182
> +#define VC_ENDC_FLUSH_INJECT 0x410
> +
> /* ENDC Sync done */
> #define X_VC_ENDC_SYNC_DONE 0x184
> #define VC_ENDC_SYNC_DONE 0x420
> @@ -403,6 +415,10 @@
> #define PC_NXC_FLUSH_POLL_BLOCK_ID_MASK PPC_BITMASK(36, 39)
> #define PC_NXC_FLUSH_POLL_OFFSET_MASK PPC_BITMASK(40, 63) /* 24-bit */
>
> +/* NxC Cache flush inject */
> +#define X_PC_NXC_FLUSH_INJECT 0x282
> +#define PC_NXC_FLUSH_INJECT 0x410
> +
> /* NxC Cache watch assign */
> #define X_PC_NXC_WATCH_ASSIGN 0x286
> #define PC_NXC_WATCH_ASSIGN 0x430
> diff --git a/include/hw/ppc/pnv_chip.h b/include/hw/ppc/pnv_chip.h
> index a4ed17ac59..9502248561 100644
> --- a/include/hw/ppc/pnv_chip.h
> +++ b/include/hw/ppc/pnv_chip.h
> @@ -131,6 +131,7 @@ struct Pnv10Chip {
>
> #define PNV10_PIR2FUSEDCORE(pir) (((pir) >> 3) & 0xf)
> #define PNV10_PIR2CHIP(pir) (((pir) >> 8) & 0x7f)
> +#define PNV10_PIR2THREAD(pir) (((pir) & 0x7f))
>
> struct PnvChipClass {
> /*< private >*/
> diff --git a/hw/intc/pnv_xive2.c b/hw/intc/pnv_xive2.c
> index 5b075a220c..d62ac20d98 100644
> --- a/hw/intc/pnv_xive2.c
> +++ b/hw/intc/pnv_xive2.c
> @@ -25,6 +25,7 @@
> #include "hw/ppc/ppc.h"
> #include "hw/qdev-properties.h"
> #include "sysemu/reset.h"
> +#include "sysemu/qtest.h"
>
> #include <libfdt.h>
>
> @@ -32,6 +33,16 @@
>
> #undef XIVE2_DEBUG
>
> +/* XIVE Sync or Flush Notification Block */
> +typedef struct XiveSfnBlock {
> + uint8_t bytes[32];
> +} XiveSfnBlock;
> +
> +/* XIVE Thread Sync or Flush Notification Area */
> +typedef struct XiveThreadNA {
> + XiveSfnBlock topo[16];
> +} XiveThreadNA;
> +
> /*
> * Virtual structures table (VST)
> */
> @@ -54,7 +65,7 @@ static const XiveVstInfo vst_infos[] = {
> [VST_NVC] = { "NVCT", sizeof(Xive2Nvgc), 16 },
>
> [VST_IC] = { "IC", 1, /* ? */ 16 }, /* Topology # */
> - [VST_SYNC] = { "SYNC", 1, /* ? */ 16 }, /* Topology # */
> + [VST_SYNC] = { "SYNC", sizeof(XiveThreadNA), 16 }, /* Topology # */
>
> /*
> * This table contains the backing store pages for the interrupt
> @@ -329,6 +340,70 @@ static int pnv_xive2_write_end(Xive2Router *xrtr, uint8_t blk, uint32_t idx,
> word_number);
> }
>
> +static inline int pnv_xive2_get_current_pir(PnvXive2 *xive)
> +{
> + if (!qtest_enabled()) {
> + PowerPCCPU *cpu = POWERPC_CPU(current_cpu);
> + return ppc_cpu_pir(cpu);
> + }
> + return 0;
> +}
> +
> +/*
> + * After SW injects a Queue Sync or Cache Flush operation, HW will notify
> + * SW of the completion of the operation by writing a byte of all 1's (0xff)
> + * to a specific memory location. The memory location is calculated by first
> + * looking up a base address in the SYNC VSD using the Topology ID of the
> + * originating thread as the "block" number. This points to a
> + * 64k block of memory that is further divided into 128 512 byte chunks of
> + * memory, which is indexed by the thread id of the requesting thread.
> + * Finally, this 512 byte chunk of memory is divided into 16 32 byte
> + * chunks which are indexed by the topology id of the targeted IC's chip.
> + * The values below are the offsets into that 32 byte chunk of memory for
> + * each type of cache flush or queue sync operation.
> + */
> +#define PNV_XIVE2_QUEUE_IPI 0x00
> +#define PNV_XIVE2_QUEUE_HW 0x01
> +#define PNV_XIVE2_QUEUE_NXC 0x02
> +#define PNV_XIVE2_QUEUE_INT 0x03
> +#define PNV_XIVE2_QUEUE_OS 0x04
> +#define PNV_XIVE2_QUEUE_POOL 0x05
> +#define PNV_XIVE2_QUEUE_HARD 0x06
> +#define PNV_XIVE2_CACHE_ENDC 0x08
> +#define PNV_XIVE2_CACHE_ESBC 0x09
> +#define PNV_XIVE2_CACHE_EASC 0x0a
> +#define PNV_XIVE2_QUEUE_NXC_LD_LCL_NCO 0x10
> +#define PNV_XIVE2_QUEUE_NXC_LD_LCL_CO 0x11
> +#define PNV_XIVE2_QUEUE_NXC_ST_LCL_NCI 0x12
> +#define PNV_XIVE2_QUEUE_NXC_ST_LCL_CI 0x13
> +#define PNV_XIVE2_QUEUE_NXC_ST_RMT_NCI 0x14
> +#define PNV_XIVE2_QUEUE_NXC_ST_RMT_CI 0x15
> +#define PNV_XIVE2_CACHE_NXC 0x18
> +
> +static int pnv_xive2_inject_notify(PnvXive2 *xive, int type)
> +{
> + uint64_t addr;
> + int pir = pnv_xive2_get_current_pir(xive);
> + int thread_nr = PNV10_PIR2THREAD(pir);
> + int thread_topo_id = PNV10_PIR2CHIP(pir);
> + int ic_topo_id = xive->chip->chip_id;
> + uint64_t offset = ic_topo_id * sizeof(XiveSfnBlock);
> + uint8_t byte = 0xff;
> +
> + /* Retrieve the address of requesting thread's notification area */
> + addr = pnv_xive2_vst_addr(xive, VST_SYNC, thread_topo_id, thread_nr);
> +
> + if (!addr) {
> + xive2_error(xive, "VST: no SYNC entry %x/%x !?",
> + thread_topo_id, thread_nr);
> + return -1;
> + }
> +
> + cpu_physical_memory_write(addr + offset + type, &byte, 1);
May be use address_space_stb() to test the result of the mem transaction ?
Thanks,
C.
> +
> + return 0;
> +}
> +
> static int pnv_xive2_end_update(PnvXive2 *xive, uint8_t watch_engine)
> {
> uint8_t blk;
> @@ -1180,6 +1255,10 @@ static void pnv_xive2_ic_vc_write(void *opaque, hwaddr offset,
> /* ESB update */
> break;
>
> + case VC_ESBC_FLUSH_INJECT:
> + pnv_xive2_inject_notify(xive, PNV_XIVE2_CACHE_ESBC);
> + break;
> +
> case VC_ESBC_CFG:
> break;
>
> @@ -1192,6 +1271,10 @@ static void pnv_xive2_ic_vc_write(void *opaque, hwaddr offset,
> /* EAS update */
> break;
>
> + case VC_EASC_FLUSH_INJECT:
> + pnv_xive2_inject_notify(xive, PNV_XIVE2_CACHE_EASC);
> + break;
> +
> case VC_ENDC_CFG:
> break;
>
> @@ -1226,6 +1309,10 @@ static void pnv_xive2_ic_vc_write(void *opaque, hwaddr offset,
> xive->vc_regs[VC_ENDC_FLUSH_CTRL >> 3] |= VC_ENDC_FLUSH_CTRL_POLL_VALID;
> break;
>
> + case VC_ENDC_FLUSH_INJECT:
> + pnv_xive2_inject_notify(xive, PNV_XIVE2_CACHE_ENDC);
> + break;
> +
> /*
> * Indirect invalidation
> */
> @@ -1426,6 +1513,10 @@ static void pnv_xive2_ic_pc_write(void *opaque, hwaddr offset,
> xive->pc_regs[PC_NXC_FLUSH_CTRL >> 3] |= PC_NXC_FLUSH_CTRL_POLL_VALID;
> break;
>
> + case PC_NXC_FLUSH_INJECT:
> + pnv_xive2_inject_notify(xive, PNV_XIVE2_CACHE_NXC);
> + break;
> +
> /*
> * Indirect invalidation
> */
> @@ -1729,6 +1820,12 @@ static const MemoryRegionOps pnv_xive2_ic_lsi_ops = {
> #define PNV_XIVE2_SYNC_OS_ESC 0x200
> #define PNV_XIVE2_SYNC_POOL_ESC 0x280
> #define PNV_XIVE2_SYNC_HARD_ESC 0x300
> +#define PNV_XIVE2_SYNC_NXC_LD_LCL_NCO 0x800
> +#define PNV_XIVE2_SYNC_NXC_LD_LCL_CO 0x880
> +#define PNV_XIVE2_SYNC_NXC_ST_LCL_NCI 0x900
> +#define PNV_XIVE2_SYNC_NXC_ST_LCL_CI 0x980
> +#define PNV_XIVE2_SYNC_NXC_ST_RMT_NCI 0xA00
> +#define PNV_XIVE2_SYNC_NXC_ST_RMT_CI 0xA80
>
> static uint64_t pnv_xive2_ic_sync_read(void *opaque, hwaddr offset,
> unsigned size)
> @@ -1740,22 +1837,72 @@ static uint64_t pnv_xive2_ic_sync_read(void *opaque, hwaddr offset,
> return -1;
> }
>
> +/*
> + * The sync MMIO space spans two pages. The lower page is use for
> + * queue sync "poll" requests while the upper page is used for queue
> + * sync "inject" requests. Inject requests require the HW to write
> + * a byte of all 1's to a predetermined location in memory in order
> + * to signal completion of the request. Both pages have the same
> + * layout, so it is easiest to handle both with a single function.
> + */
> static void pnv_xive2_ic_sync_write(void *opaque, hwaddr offset,
> uint64_t val, unsigned size)
> {
> PnvXive2 *xive = PNV_XIVE2(opaque);
> + int inject_type;
> + hwaddr pg_offset_mask = (1ull << xive->ic_shift) - 1;
>
> - switch (offset) {
> + /* adjust offset for inject page */
> + hwaddr adj_offset = offset & pg_offset_mask;
> +
> + switch (adj_offset) {
> case PNV_XIVE2_SYNC_IPI:
> + inject_type = PNV_XIVE2_QUEUE_IPI;
> + break;
> case PNV_XIVE2_SYNC_HW:
> + inject_type = PNV_XIVE2_QUEUE_HW;
> + break;
> case PNV_XIVE2_SYNC_NxC:
> + inject_type = PNV_XIVE2_QUEUE_NXC;
> + break;
> case PNV_XIVE2_SYNC_INT:
> + inject_type = PNV_XIVE2_QUEUE_INT;
> + break;
> case PNV_XIVE2_SYNC_OS_ESC:
> + inject_type = PNV_XIVE2_QUEUE_OS;
> + break;
> case PNV_XIVE2_SYNC_POOL_ESC:
> + inject_type = PNV_XIVE2_QUEUE_POOL;
> + break;
> case PNV_XIVE2_SYNC_HARD_ESC:
> + inject_type = PNV_XIVE2_QUEUE_HARD;
> + break;
> + case PNV_XIVE2_SYNC_NXC_LD_LCL_NCO:
> + inject_type = PNV_XIVE2_QUEUE_NXC_LD_LCL_NCO;
> + break;
> + case PNV_XIVE2_SYNC_NXC_LD_LCL_CO:
> + inject_type = PNV_XIVE2_QUEUE_NXC_LD_LCL_CO;
> + break;
> + case PNV_XIVE2_SYNC_NXC_ST_LCL_NCI:
> + inject_type = PNV_XIVE2_QUEUE_NXC_ST_LCL_NCI;
> + break;
> + case PNV_XIVE2_SYNC_NXC_ST_LCL_CI:
> + inject_type = PNV_XIVE2_QUEUE_NXC_ST_LCL_CI;
> + break;
> + case PNV_XIVE2_SYNC_NXC_ST_RMT_NCI:
> + inject_type = PNV_XIVE2_QUEUE_NXC_ST_RMT_NCI;
> + break;
> + case PNV_XIVE2_SYNC_NXC_ST_RMT_CI:
> + inject_type = PNV_XIVE2_QUEUE_NXC_ST_RMT_CI;
> break;
> default:
> xive2_error(xive, "SYNC: invalid write @%"HWADDR_PRIx, offset);
> + return;
> + }
> +
> + /* Write Queue Sync notification byte if writing to sync inject page */
> + if ((offset & ~pg_offset_mask) != 0) {
> + pnv_xive2_inject_notify(xive, inject_type);
> }
> }
>
^ permalink raw reply [flat|nested] 23+ messages in thread
* [PATCH v2 4/9] pnv/xive2: Add NVG and NVC to cache watch facility
2024-07-15 18:33 [PATCH v2 0/9] XIVE changes for Cache Watch, VSTs, STT and info pic Michael Kowal
` (2 preceding siblings ...)
2024-07-15 18:33 ` [PATCH v2 3/9] pnv/xive: Support cache flush and queue sync inject with notifications Michael Kowal
@ 2024-07-15 18:33 ` Michael Kowal
2024-07-16 7:29 ` Cédric Le Goater
2024-07-15 18:33 ` [PATCH v2 5/9] pnv/xive2: Configure Virtualization Structure Tables through the PC Michael Kowal
` (4 subsequent siblings)
8 siblings, 1 reply; 23+ messages in thread
From: Michael Kowal @ 2024-07-15 18:33 UTC (permalink / raw)
To: qemu-devel; +Cc: qemu-ppc, clg, fbarrat, npiggin, milesg
From: Frederic Barrat <fbarrat@linux.ibm.com>
The cache watch facility uses the same register interface to handle
entries in the NVP, NVG and NVC tables. A bit-field in the 'watchX
specification' register tells the table type. So far, that bit-field
was not read and the code assumed a read/write to the NVP table.
This patch allows to read/write entries in the NVG and NVC table as
well.
Signed-off-by: Frederic Barrat <fbarrat@linux.ibm.com>
Signed-off-by: Michael Kowal <kowal@linux.vnet.ibm.com>
---
hw/intc/pnv_xive2.c | 59 ++++++++++++++++++++++++++++++++-------------
1 file changed, 42 insertions(+), 17 deletions(-)
diff --git a/hw/intc/pnv_xive2.c b/hw/intc/pnv_xive2.c
index d62ac20d98..c72c66dd6a 100644
--- a/hw/intc/pnv_xive2.c
+++ b/hw/intc/pnv_xive2.c
@@ -462,46 +462,71 @@ static int pnv_xive2_write_nvp(Xive2Router *xrtr, uint8_t blk, uint32_t idx,
word_number);
}
-static int pnv_xive2_nvp_update(PnvXive2 *xive, uint8_t watch_engine)
+static int pnv_xive2_nxc_to_table_type(uint8_t nxc_type, uint32_t *table_type)
{
- uint8_t blk;
- uint32_t idx;
+ switch (nxc_type) {
+ case PC_NXC_WATCH_NXC_NVP:
+ *table_type = VST_NVP;
+ break;
+ case PC_NXC_WATCH_NXC_NVG:
+ *table_type = VST_NVG;
+ break;
+ case PC_NXC_WATCH_NXC_NVC:
+ *table_type = VST_NVC;
+ break;
+ default:
+ qemu_log_mask(LOG_GUEST_ERROR,
+ "XIVE: invalid table type for nxc operation\n");
+ return -1;
+ }
+ return 0;
+}
+
+static int pnv_xive2_nxc_update(PnvXive2 *xive, uint8_t watch_engine)
+{
+ uint8_t blk, nxc_type;
+ uint32_t idx, table_type = -1;
int i, spec_reg, data_reg;
uint64_t nxc_watch[4];
- if (watch_engine > 3) {
- return -1;
- }
+ assert(watch_engine < ARRAY_SIZE(nxc_watch));
+
spec_reg = (PC_NXC_WATCH0_SPEC + watch_engine * 0x40) >> 3;
data_reg = (PC_NXC_WATCH0_DATA0 + watch_engine * 0x40) >> 3;
+ nxc_type = GETFIELD(PC_NXC_WATCH_NXC_TYPE, xive->pc_regs[spec_reg]);
blk = GETFIELD(PC_NXC_WATCH_BLOCK_ID, xive->pc_regs[spec_reg]);
idx = GETFIELD(PC_NXC_WATCH_INDEX, xive->pc_regs[spec_reg]);
+ assert(pnv_xive2_nxc_to_table_type(nxc_type, &table_type));
+
for (i = 0; i < ARRAY_SIZE(nxc_watch); i++) {
nxc_watch[i] = cpu_to_be64(xive->pc_regs[data_reg + i]);
}
- return pnv_xive2_vst_write(xive, VST_NVP, blk, idx, nxc_watch,
+ return pnv_xive2_vst_write(xive, table_type, blk, idx, nxc_watch,
XIVE_VST_WORD_ALL);
}
-static void pnv_xive2_nvp_cache_load(PnvXive2 *xive, uint8_t watch_engine)
+static void pnv_xive2_nxc_cache_load(PnvXive2 *xive, uint8_t watch_engine)
{
- uint8_t blk;
- uint32_t idx;
+ uint8_t blk, nxc_type;
+ uint32_t idx, table_type = -1;
uint64_t nxc_watch[4] = { 0 };
int i, spec_reg, data_reg;
- if (watch_engine > 3) {
- return;
- }
+ assert(watch_engine < ARRAY_SIZE(nxc_watch));
+
spec_reg = (PC_NXC_WATCH0_SPEC + watch_engine * 0x40) >> 3;
data_reg = (PC_NXC_WATCH0_DATA0 + watch_engine * 0x40) >> 3;
+ nxc_type = GETFIELD(PC_NXC_WATCH_NXC_TYPE, xive->pc_regs[spec_reg]);
blk = GETFIELD(PC_NXC_WATCH_BLOCK_ID, xive->pc_regs[spec_reg]);
idx = GETFIELD(PC_NXC_WATCH_INDEX, xive->pc_regs[spec_reg]);
- if (pnv_xive2_vst_read(xive, VST_NVP, blk, idx, nxc_watch)) {
- xive2_error(xive, "VST: no NVP entry %x/%x !?", blk, idx);
+ assert(pnv_xive2_nxc_to_table_type(nxc_type, &table_type));
+
+ if (pnv_xive2_vst_read(xive, table_type, blk, idx, nxc_watch)) {
+ xive2_error(xive, "VST: no NXC entry %x/%x in %s table!?",
+ blk, idx, vst_infos[table_type].name);
}
for (i = 0; i < ARRAY_SIZE(nxc_watch); i++) {
@@ -1431,7 +1456,7 @@ static uint64_t pnv_xive2_ic_pc_read(void *opaque, hwaddr offset,
* SPEC register
*/
watch_engine = (offset - PC_NXC_WATCH0_DATA0) >> 6;
- pnv_xive2_nvp_cache_load(xive, watch_engine);
+ pnv_xive2_nxc_cache_load(xive, watch_engine);
val = xive->pc_regs[reg];
break;
@@ -1505,7 +1530,7 @@ static void pnv_xive2_ic_pc_write(void *opaque, hwaddr offset,
/* writing to DATA0 triggers the cache write */
watch_engine = (offset - PC_NXC_WATCH0_DATA0) >> 6;
xive->pc_regs[reg] = val;
- pnv_xive2_nvp_update(xive, watch_engine);
+ pnv_xive2_nxc_update(xive, watch_engine);
break;
/* case PC_NXC_FLUSH_CTRL: */
--
2.43.0
^ permalink raw reply related [flat|nested] 23+ messages in thread* Re: [PATCH v2 4/9] pnv/xive2: Add NVG and NVC to cache watch facility
2024-07-15 18:33 ` [PATCH v2 4/9] pnv/xive2: Add NVG and NVC to cache watch facility Michael Kowal
@ 2024-07-16 7:29 ` Cédric Le Goater
2024-07-16 15:46 ` Mike Kowal
0 siblings, 1 reply; 23+ messages in thread
From: Cédric Le Goater @ 2024-07-16 7:29 UTC (permalink / raw)
To: Michael Kowal, qemu-devel; +Cc: qemu-ppc, fbarrat, npiggin, milesg
On 7/15/24 20:33, Michael Kowal wrote:
> From: Frederic Barrat <fbarrat@linux.ibm.com>
>
> The cache watch facility uses the same register interface to handle
> entries in the NVP, NVG and NVC tables. A bit-field in the 'watchX
> specification' register tells the table type. So far, that bit-field
> was not read and the code assumed a read/write to the NVP table.
>
> This patch allows to read/write entries in the NVG and NVC table as
> well.
>
> Signed-off-by: Frederic Barrat <fbarrat@linux.ibm.com>
> Signed-off-by: Michael Kowal <kowal@linux.vnet.ibm.com>
> ---
> hw/intc/pnv_xive2.c | 59 ++++++++++++++++++++++++++++++++-------------
> 1 file changed, 42 insertions(+), 17 deletions(-)
>
> diff --git a/hw/intc/pnv_xive2.c b/hw/intc/pnv_xive2.c
> index d62ac20d98..c72c66dd6a 100644
> --- a/hw/intc/pnv_xive2.c
> +++ b/hw/intc/pnv_xive2.c
> @@ -462,46 +462,71 @@ static int pnv_xive2_write_nvp(Xive2Router *xrtr, uint8_t blk, uint32_t idx,
> word_number);
> }
>
> -static int pnv_xive2_nvp_update(PnvXive2 *xive, uint8_t watch_engine)
> +static int pnv_xive2_nxc_to_table_type(uint8_t nxc_type, uint32_t *table_type)
> {
> - uint8_t blk;
> - uint32_t idx;
> + switch (nxc_type) {
> + case PC_NXC_WATCH_NXC_NVP:
> + *table_type = VST_NVP;
> + break;
> + case PC_NXC_WATCH_NXC_NVG:
> + *table_type = VST_NVG;
> + break;
> + case PC_NXC_WATCH_NXC_NVC:
> + *table_type = VST_NVC;
> + break;
> + default:
> + qemu_log_mask(LOG_GUEST_ERROR,
> + "XIVE: invalid table type for nxc operation\n");
> + return -1;
> + }
> + return 0;
> +}
> +
> +static int pnv_xive2_nxc_update(PnvXive2 *xive, uint8_t watch_engine)
> +{
> + uint8_t blk, nxc_type;
> + uint32_t idx, table_type = -1;
> int i, spec_reg, data_reg;
> uint64_t nxc_watch[4];
>
> - if (watch_engine > 3) {
> - return -1;
> - }
> + assert(watch_engine < ARRAY_SIZE(nxc_watch));
> +
This belongs to patch 1.
> spec_reg = (PC_NXC_WATCH0_SPEC + watch_engine * 0x40) >> 3;
> data_reg = (PC_NXC_WATCH0_DATA0 + watch_engine * 0x40) >> 3;
> + nxc_type = GETFIELD(PC_NXC_WATCH_NXC_TYPE, xive->pc_regs[spec_reg]);
> blk = GETFIELD(PC_NXC_WATCH_BLOCK_ID, xive->pc_regs[spec_reg]);
> idx = GETFIELD(PC_NXC_WATCH_INDEX, xive->pc_regs[spec_reg]);
>
> + assert(pnv_xive2_nxc_to_table_type(nxc_type, &table_type));
> +
> for (i = 0; i < ARRAY_SIZE(nxc_watch); i++) {
> nxc_watch[i] = cpu_to_be64(xive->pc_regs[data_reg + i]);
> }
>
> - return pnv_xive2_vst_write(xive, VST_NVP, blk, idx, nxc_watch,
> + return pnv_xive2_vst_write(xive, table_type, blk, idx, nxc_watch,
> XIVE_VST_WORD_ALL);
> }
>
> -static void pnv_xive2_nvp_cache_load(PnvXive2 *xive, uint8_t watch_engine)
> +static void pnv_xive2_nxc_cache_load(PnvXive2 *xive, uint8_t watch_engine)
> {
> - uint8_t blk;
> - uint32_t idx;
> + uint8_t blk, nxc_type;
> + uint32_t idx, table_type = -1;
> uint64_t nxc_watch[4] = { 0 };
> int i, spec_reg, data_reg;
>
> - if (watch_engine > 3) {
> - return;
> - }
> + assert(watch_engine < ARRAY_SIZE(nxc_watch));
> +
> spec_reg = (PC_NXC_WATCH0_SPEC + watch_engine * 0x40) >> 3;
> data_reg = (PC_NXC_WATCH0_DATA0 + watch_engine * 0x40) >> 3;
> + nxc_type = GETFIELD(PC_NXC_WATCH_NXC_TYPE, xive->pc_regs[spec_reg]);
> blk = GETFIELD(PC_NXC_WATCH_BLOCK_ID, xive->pc_regs[spec_reg]);
> idx = GETFIELD(PC_NXC_WATCH_INDEX, xive->pc_regs[spec_reg]);
>
> - if (pnv_xive2_vst_read(xive, VST_NVP, blk, idx, nxc_watch)) {
> - xive2_error(xive, "VST: no NVP entry %x/%x !?", blk, idx);
> + assert(pnv_xive2_nxc_to_table_type(nxc_type, &table_type));
> +
> + if (pnv_xive2_vst_read(xive, table_type, blk, idx, nxc_watch)) {
> + xive2_error(xive, "VST: no NXC entry %x/%x in %s table!?",
> + blk, idx, vst_infos[table_type].name);
> }
>
> for (i = 0; i < ARRAY_SIZE(nxc_watch); i++) {
> @@ -1431,7 +1456,7 @@ static uint64_t pnv_xive2_ic_pc_read(void *opaque, hwaddr offset,
> * SPEC register
> */
> watch_engine = (offset - PC_NXC_WATCH0_DATA0) >> 6;
> - pnv_xive2_nvp_cache_load(xive, watch_engine);
> + pnv_xive2_nxc_cache_load(xive, watch_engine);
> val = xive->pc_regs[reg];
> break;
>
> @@ -1505,7 +1530,7 @@ static void pnv_xive2_ic_pc_write(void *opaque, hwaddr offset,
> /* writing to DATA0 triggers the cache write */
> watch_engine = (offset - PC_NXC_WATCH0_DATA0) >> 6;
> xive->pc_regs[reg] = val;
> - pnv_xive2_nvp_update(xive, watch_engine);
> + pnv_xive2_nxc_update(xive, watch_engine);
> break;
>
> /* case PC_NXC_FLUSH_CTRL: */
^ permalink raw reply [flat|nested] 23+ messages in thread* Re: [PATCH v2 4/9] pnv/xive2: Add NVG and NVC to cache watch facility
2024-07-16 7:29 ` Cédric Le Goater
@ 2024-07-16 15:46 ` Mike Kowal
2024-07-16 18:39 ` Mike Kowal
0 siblings, 1 reply; 23+ messages in thread
From: Mike Kowal @ 2024-07-16 15:46 UTC (permalink / raw)
To: Cédric Le Goater, Michael Kowal, qemu-devel
Cc: qemu-ppc, fbarrat, npiggin, milesg
On 7/16/2024 2:29 AM, Cédric Le Goater wrote:
> On 7/15/24 20:33, Michael Kowal wrote:
>> From: Frederic Barrat <fbarrat@linux.ibm.com>
>>
>> The cache watch facility uses the same register interface to handle
>> entries in the NVP, NVG and NVC tables. A bit-field in the 'watchX
>> specification' register tells the table type. So far, that bit-field
>> was not read and the code assumed a read/write to the NVP table.
>>
>> This patch allows to read/write entries in the NVG and NVC table as
>> well.
>>
>> Signed-off-by: Frederic Barrat <fbarrat@linux.ibm.com>
>> Signed-off-by: Michael Kowal <kowal@linux.vnet.ibm.com>
>> ---
>> hw/intc/pnv_xive2.c | 59 ++++++++++++++++++++++++++++++++-------------
>> 1 file changed, 42 insertions(+), 17 deletions(-)
>>
>> diff --git a/hw/intc/pnv_xive2.c b/hw/intc/pnv_xive2.c
>> index d62ac20d98..c72c66dd6a 100644
>> --- a/hw/intc/pnv_xive2.c
>> +++ b/hw/intc/pnv_xive2.c
>> @@ -462,46 +462,71 @@ static int pnv_xive2_write_nvp(Xive2Router
>> *xrtr, uint8_t blk, uint32_t idx,
>> word_number);
>> }
>> -static int pnv_xive2_nvp_update(PnvXive2 *xive, uint8_t watch_engine)
>> +static int pnv_xive2_nxc_to_table_type(uint8_t nxc_type, uint32_t
>> *table_type)
>> {
>> - uint8_t blk;
>> - uint32_t idx;
>> + switch (nxc_type) {
>> + case PC_NXC_WATCH_NXC_NVP:
>> + *table_type = VST_NVP;
>> + break;
>> + case PC_NXC_WATCH_NXC_NVG:
>> + *table_type = VST_NVG;
>> + break;
>> + case PC_NXC_WATCH_NXC_NVC:
>> + *table_type = VST_NVC;
>> + break;
>> + default:
>> + qemu_log_mask(LOG_GUEST_ERROR,
>> + "XIVE: invalid table type for nxc operation\n");
>> + return -1;
>> + }
>> + return 0;
>> +}
>> +
>> +static int pnv_xive2_nxc_update(PnvXive2 *xive, uint8_t watch_engine)
>> +{
>> + uint8_t blk, nxc_type;
>> + uint32_t idx, table_type = -1;
>> int i, spec_reg, data_reg;
>> uint64_t nxc_watch[4];
>> - if (watch_engine > 3) {
>> - return -1;
>> - }
>> + assert(watch_engine < ARRAY_SIZE(nxc_watch));
>> +
>
> This belongs to patch 1.
I was trying to reduce the size of the first patch set and this could be
isolated to separate patch set. I can move it back to patch set 1.
Also note above, this is where I originally added the assert...
>
>> spec_reg = (PC_NXC_WATCH0_SPEC + watch_engine * 0x40) >> 3;
>> data_reg = (PC_NXC_WATCH0_DATA0 + watch_engine * 0x40) >> 3;
>> + nxc_type = GETFIELD(PC_NXC_WATCH_NXC_TYPE,
>> xive->pc_regs[spec_reg]);
>> blk = GETFIELD(PC_NXC_WATCH_BLOCK_ID, xive->pc_regs[spec_reg]);
>> idx = GETFIELD(PC_NXC_WATCH_INDEX, xive->pc_regs[spec_reg]);
>> + assert(pnv_xive2_nxc_to_table_type(nxc_type, &table_type));
>> +
>> for (i = 0; i < ARRAY_SIZE(nxc_watch); i++) {
>> nxc_watch[i] = cpu_to_be64(xive->pc_regs[data_reg + i]);
>> }
>> - return pnv_xive2_vst_write(xive, VST_NVP, blk, idx, nxc_watch,
>> + return pnv_xive2_vst_write(xive, table_type, blk, idx, nxc_watch,
>> XIVE_VST_WORD_ALL);
>> }
>> -static void pnv_xive2_nvp_cache_load(PnvXive2 *xive, uint8_t
>> watch_engine)
>> +static void pnv_xive2_nxc_cache_load(PnvXive2 *xive, uint8_t
>> watch_engine)
>> {
>> - uint8_t blk;
>> - uint32_t idx;
>> + uint8_t blk, nxc_type;
>> + uint32_t idx, table_type = -1;
>> uint64_t nxc_watch[4] = { 0 };
>> int i, spec_reg, data_reg;
>> - if (watch_engine > 3) {
>> - return;
>> - }
>> + assert(watch_engine < ARRAY_SIZE(nxc_watch));
>> +
>> spec_reg = (PC_NXC_WATCH0_SPEC + watch_engine * 0x40) >> 3;
>> data_reg = (PC_NXC_WATCH0_DATA0 + watch_engine * 0x40) >> 3;
>> + nxc_type = GETFIELD(PC_NXC_WATCH_NXC_TYPE,
>> xive->pc_regs[spec_reg]);
>> blk = GETFIELD(PC_NXC_WATCH_BLOCK_ID, xive->pc_regs[spec_reg]);
>> idx = GETFIELD(PC_NXC_WATCH_INDEX, xive->pc_regs[spec_reg]);
>> - if (pnv_xive2_vst_read(xive, VST_NVP, blk, idx, nxc_watch)) {
>> - xive2_error(xive, "VST: no NVP entry %x/%x !?", blk, idx);
>> + assert(pnv_xive2_nxc_to_table_type(nxc_type, &table_type));
>> +
>> + if (pnv_xive2_vst_read(xive, table_type, blk, idx, nxc_watch)) {
>> + xive2_error(xive, "VST: no NXC entry %x/%x in %s table!?",
>> + blk, idx, vst_infos[table_type].name);
>> }
>> for (i = 0; i < ARRAY_SIZE(nxc_watch); i++) {
>> @@ -1431,7 +1456,7 @@ static uint64_t pnv_xive2_ic_pc_read(void
>> *opaque, hwaddr offset,
>> * SPEC register
>> */
>> watch_engine = (offset - PC_NXC_WATCH0_DATA0) >> 6;
>> - pnv_xive2_nvp_cache_load(xive, watch_engine);
>> + pnv_xive2_nxc_cache_load(xive, watch_engine);
>> val = xive->pc_regs[reg];
>> break;
>> @@ -1505,7 +1530,7 @@ static void pnv_xive2_ic_pc_write(void
>> *opaque, hwaddr offset,
>> /* writing to DATA0 triggers the cache write */
>> watch_engine = (offset - PC_NXC_WATCH0_DATA0) >> 6;
>> xive->pc_regs[reg] = val;
>> - pnv_xive2_nvp_update(xive, watch_engine);
>> + pnv_xive2_nxc_update(xive, watch_engine);
>> break;
>> /* case PC_NXC_FLUSH_CTRL: */
>
^ permalink raw reply [flat|nested] 23+ messages in thread* Re: [PATCH v2 4/9] pnv/xive2: Add NVG and NVC to cache watch facility
2024-07-16 15:46 ` Mike Kowal
@ 2024-07-16 18:39 ` Mike Kowal
0 siblings, 0 replies; 23+ messages in thread
From: Mike Kowal @ 2024-07-16 18:39 UTC (permalink / raw)
To: Cédric Le Goater, Michael Kowal, qemu-devel
Cc: qemu-ppc, fbarrat, npiggin, milesg
On 7/16/2024 10:46 AM, Mike Kowal wrote:
>
> On 7/16/2024 2:29 AM, Cédric Le Goater wrote:
>> On 7/15/24 20:33, Michael Kowal wrote:
>>> From: Frederic Barrat <fbarrat@linux.ibm.com>
>>>
>>> The cache watch facility uses the same register interface to handle
>>> entries in the NVP, NVG and NVC tables. A bit-field in the 'watchX
>>> specification' register tells the table type. So far, that bit-field
>>> was not read and the code assumed a read/write to the NVP table.
>>>
>>> This patch allows to read/write entries in the NVG and NVC table as
>>> well.
>>>
>>> Signed-off-by: Frederic Barrat <fbarrat@linux.ibm.com>
>>> Signed-off-by: Michael Kowal <kowal@linux.vnet.ibm.com>
>>> ---
>>> hw/intc/pnv_xive2.c | 59
>>> ++++++++++++++++++++++++++++++++-------------
>>> 1 file changed, 42 insertions(+), 17 deletions(-)
>>>
>>> diff --git a/hw/intc/pnv_xive2.c b/hw/intc/pnv_xive2.c
>>> index d62ac20d98..c72c66dd6a 100644
>>> --- a/hw/intc/pnv_xive2.c
>>> +++ b/hw/intc/pnv_xive2.c
>>> @@ -462,46 +462,71 @@ static int pnv_xive2_write_nvp(Xive2Router
>>> *xrtr, uint8_t blk, uint32_t idx,
>>> word_number);
>>> }
>>> -static int pnv_xive2_nvp_update(PnvXive2 *xive, uint8_t
>>> watch_engine)
>>> +static int pnv_xive2_nxc_to_table_type(uint8_t nxc_type, uint32_t
>>> *table_type)
>>> {
>>> - uint8_t blk;
>>> - uint32_t idx;
>>> + switch (nxc_type) {
>>> + case PC_NXC_WATCH_NXC_NVP:
>>> + *table_type = VST_NVP;
>>> + break;
>>> + case PC_NXC_WATCH_NXC_NVG:
>>> + *table_type = VST_NVG;
>>> + break;
>>> + case PC_NXC_WATCH_NXC_NVC:
>>> + *table_type = VST_NVC;
>>> + break;
>>> + default:
>>> + qemu_log_mask(LOG_GUEST_ERROR,
>>> + "XIVE: invalid table type for nxc operation\n");
>>> + return -1;
>>> + }
>>> + return 0;
>>> +}
>>> +
>>> +static int pnv_xive2_nxc_update(PnvXive2 *xive, uint8_t watch_engine)
>>> +{
>>> + uint8_t blk, nxc_type;
>>> + uint32_t idx, table_type = -1;
>>> int i, spec_reg, data_reg;
>>> uint64_t nxc_watch[4];
>>> - if (watch_engine > 3) {
>>> - return -1;
>>> - }
>>> + assert(watch_engine < ARRAY_SIZE(nxc_watch));
>>> +
>>
>> This belongs to patch 1.
>
> I was trying to reduce the size of the first patch set and this could
> be isolated to separate patch set. I can move it back to patch set 1.
>
> Also note above, this is where I originally added the assert...
Disregard these comments. After rereading I understand. I thought you
meant that this entire patch set should be in 1...
>
>>
>>> spec_reg = (PC_NXC_WATCH0_SPEC + watch_engine * 0x40) >> 3;
>>> data_reg = (PC_NXC_WATCH0_DATA0 + watch_engine * 0x40) >> 3;
>>> + nxc_type = GETFIELD(PC_NXC_WATCH_NXC_TYPE,
>>> xive->pc_regs[spec_reg]);
>>> blk = GETFIELD(PC_NXC_WATCH_BLOCK_ID, xive->pc_regs[spec_reg]);
>>> idx = GETFIELD(PC_NXC_WATCH_INDEX, xive->pc_regs[spec_reg]);
>>> + assert(pnv_xive2_nxc_to_table_type(nxc_type, &table_type));
>>> +
>>> for (i = 0; i < ARRAY_SIZE(nxc_watch); i++) {
>>> nxc_watch[i] = cpu_to_be64(xive->pc_regs[data_reg + i]);
>>> }
>>> - return pnv_xive2_vst_write(xive, VST_NVP, blk, idx, nxc_watch,
>>> + return pnv_xive2_vst_write(xive, table_type, blk, idx, nxc_watch,
>>> XIVE_VST_WORD_ALL);
>>> }
>>> -static void pnv_xive2_nvp_cache_load(PnvXive2 *xive, uint8_t
>>> watch_engine)
>>> +static void pnv_xive2_nxc_cache_load(PnvXive2 *xive, uint8_t
>>> watch_engine)
>>> {
>>> - uint8_t blk;
>>> - uint32_t idx;
>>> + uint8_t blk, nxc_type;
>>> + uint32_t idx, table_type = -1;
>>> uint64_t nxc_watch[4] = { 0 };
>>> int i, spec_reg, data_reg;
>>> - if (watch_engine > 3) {
>>> - return;
>>> - }
>>> + assert(watch_engine < ARRAY_SIZE(nxc_watch));
>>> +
>>> spec_reg = (PC_NXC_WATCH0_SPEC + watch_engine * 0x40) >> 3;
>>> data_reg = (PC_NXC_WATCH0_DATA0 + watch_engine * 0x40) >> 3;
>>> + nxc_type = GETFIELD(PC_NXC_WATCH_NXC_TYPE,
>>> xive->pc_regs[spec_reg]);
>>> blk = GETFIELD(PC_NXC_WATCH_BLOCK_ID, xive->pc_regs[spec_reg]);
>>> idx = GETFIELD(PC_NXC_WATCH_INDEX, xive->pc_regs[spec_reg]);
>>> - if (pnv_xive2_vst_read(xive, VST_NVP, blk, idx, nxc_watch)) {
>>> - xive2_error(xive, "VST: no NVP entry %x/%x !?", blk, idx);
>>> + assert(pnv_xive2_nxc_to_table_type(nxc_type, &table_type));
>>> +
>>> + if (pnv_xive2_vst_read(xive, table_type, blk, idx, nxc_watch)) {
>>> + xive2_error(xive, "VST: no NXC entry %x/%x in %s table!?",
>>> + blk, idx, vst_infos[table_type].name);
>>> }
>>> for (i = 0; i < ARRAY_SIZE(nxc_watch); i++) {
>>> @@ -1431,7 +1456,7 @@ static uint64_t pnv_xive2_ic_pc_read(void
>>> *opaque, hwaddr offset,
>>> * SPEC register
>>> */
>>> watch_engine = (offset - PC_NXC_WATCH0_DATA0) >> 6;
>>> - pnv_xive2_nvp_cache_load(xive, watch_engine);
>>> + pnv_xive2_nxc_cache_load(xive, watch_engine);
>>> val = xive->pc_regs[reg];
>>> break;
>>> @@ -1505,7 +1530,7 @@ static void pnv_xive2_ic_pc_write(void
>>> *opaque, hwaddr offset,
>>> /* writing to DATA0 triggers the cache write */
>>> watch_engine = (offset - PC_NXC_WATCH0_DATA0) >> 6;
>>> xive->pc_regs[reg] = val;
>>> - pnv_xive2_nvp_update(xive, watch_engine);
>>> + pnv_xive2_nxc_update(xive, watch_engine);
>>> break;
>>> /* case PC_NXC_FLUSH_CTRL: */
>>
^ permalink raw reply [flat|nested] 23+ messages in thread
* [PATCH v2 5/9] pnv/xive2: Configure Virtualization Structure Tables through the PC
2024-07-15 18:33 [PATCH v2 0/9] XIVE changes for Cache Watch, VSTs, STT and info pic Michael Kowal
` (3 preceding siblings ...)
2024-07-15 18:33 ` [PATCH v2 4/9] pnv/xive2: Add NVG and NVC to cache watch facility Michael Kowal
@ 2024-07-15 18:33 ` Michael Kowal
2024-07-16 7:06 ` Cédric Le Goater
2024-07-15 18:33 ` [PATCH v2 6/9] pnv/xive2: Enable VST NVG and NVC index compression Michael Kowal
` (3 subsequent siblings)
8 siblings, 1 reply; 23+ messages in thread
From: Michael Kowal @ 2024-07-15 18:33 UTC (permalink / raw)
To: qemu-devel; +Cc: qemu-ppc, clg, fbarrat, npiggin, milesg
From: Frederic Barrat <fbarrat@linux.ibm.com>
Both the virtualization layer (VC) and presentation layer (PC) need to
be configured to access the VSTs. Since the information is redundant,
the xive model combines both into one set of tables and only the
definitions going through the VC are kept. The definitions through the
PC are ignored. That works well as long as firmware calls the VC for
all the tables.
For the NVG and NVC tables, it can make sense to only configure them
with the PC, since they are only used by the presenter. So this patch
allows firmware to configure the VST tables through the PC as well.
The definitions are still shared, since the VST tables can be set
through both the VC and/or PC, they are dynamically re-mapped in
memory by first deleting the memory subregion.
Signed-off-by: Frederic Barrat <fbarrat@linux.ibm.com>
Signed-off-by: Michael Kowal <kowal@linux.vnet.ibm.com>
---
hw/intc/pnv_xive2.c | 47 ++++++++++++++++++++++++++++++++++++---------
1 file changed, 38 insertions(+), 9 deletions(-)
diff --git a/hw/intc/pnv_xive2.c b/hw/intc/pnv_xive2.c
index c72c66dd6a..9e45161869 100644
--- a/hw/intc/pnv_xive2.c
+++ b/hw/intc/pnv_xive2.c
@@ -759,6 +759,9 @@ static void pnv_xive2_vst_set_exclusive(PnvXive2 *xive, uint8_t type,
* entries provisioned by FW (such as skiboot) and resize the
* ESB window accordingly.
*/
+ if (memory_region_is_mapped(&xsrc->esb_mmio)) {
+ memory_region_del_subregion(&xive->esb_mmio, &xsrc->esb_mmio);
+ }
if (!(VSD_INDIRECT & vsd)) {
memory_region_set_size(&xsrc->esb_mmio, vst_tsize * SBE_PER_BYTE
* (1ull << xsrc->esb_shift));
@@ -774,6 +777,9 @@ static void pnv_xive2_vst_set_exclusive(PnvXive2 *xive, uint8_t type,
/*
* Backing store pages for the END.
*/
+ if (memory_region_is_mapped(&end_xsrc->esb_mmio)) {
+ memory_region_del_subregion(&xive->end_mmio, &end_xsrc->esb_mmio);
+ }
if (!(VSD_INDIRECT & vsd)) {
memory_region_set_size(&end_xsrc->esb_mmio, (vst_tsize / info->size)
* (1ull << end_xsrc->esb_shift));
@@ -798,13 +804,10 @@ static void pnv_xive2_vst_set_exclusive(PnvXive2 *xive, uint8_t type,
* Both PC and VC sub-engines are configured as each use the Virtual
* Structure Tables
*/
-static void pnv_xive2_vst_set_data(PnvXive2 *xive, uint64_t vsd)
+static void pnv_xive2_vst_set_data(PnvXive2 *xive, uint64_t vsd,
+ uint8_t type, uint8_t blk)
{
uint8_t mode = GETFIELD(VSD_MODE, vsd);
- uint8_t type = GETFIELD(VC_VSD_TABLE_SELECT,
- xive->vc_regs[VC_VSD_TABLE_ADDR >> 3]);
- uint8_t blk = GETFIELD(VC_VSD_TABLE_ADDRESS,
- xive->vc_regs[VC_VSD_TABLE_ADDR >> 3]);
uint64_t vst_addr = vsd & VSD_ADDRESS_MASK;
if (type > VST_ERQ) {
@@ -839,6 +842,16 @@ static void pnv_xive2_vst_set_data(PnvXive2 *xive, uint64_t vsd)
}
}
+static void pnv_xive2_vc_vst_set_data(PnvXive2 *xive, uint64_t vsd)
+{
+ uint8_t type = GETFIELD(VC_VSD_TABLE_SELECT,
+ xive->vc_regs[VC_VSD_TABLE_ADDR >> 3]);
+ uint8_t blk = GETFIELD(VC_VSD_TABLE_ADDRESS,
+ xive->vc_regs[VC_VSD_TABLE_ADDR >> 3]);
+
+ pnv_xive2_vst_set_data(xive, vsd, type, blk);
+}
+
/*
* MMIO handlers
*/
@@ -1268,7 +1281,7 @@ static void pnv_xive2_ic_vc_write(void *opaque, hwaddr offset,
case VC_VSD_TABLE_ADDR:
break;
case VC_VSD_TABLE_DATA:
- pnv_xive2_vst_set_data(xive, val);
+ pnv_xive2_vc_vst_set_data(xive, val);
break;
/*
@@ -1487,6 +1500,16 @@ static uint64_t pnv_xive2_ic_pc_read(void *opaque, hwaddr offset,
return val;
}
+static void pnv_xive2_pc_vst_set_data(PnvXive2 *xive, uint64_t vsd)
+{
+ uint8_t type = GETFIELD(PC_VSD_TABLE_SELECT,
+ xive->pc_regs[PC_VSD_TABLE_ADDR >> 3]);
+ uint8_t blk = GETFIELD(PC_VSD_TABLE_ADDRESS,
+ xive->pc_regs[PC_VSD_TABLE_ADDR >> 3]);
+
+ pnv_xive2_vst_set_data(xive, vsd, type, blk);
+}
+
static void pnv_xive2_ic_pc_write(void *opaque, hwaddr offset,
uint64_t val, unsigned size)
{
@@ -1497,12 +1520,18 @@ static void pnv_xive2_ic_pc_write(void *opaque, hwaddr offset,
switch (offset) {
/*
- * VSD table settings. Only taken into account in the VC
- * sub-engine because the Xive2Router model combines both VC and PC
- * sub-engines
+ * VSD table settings.
+ * The Xive2Router model combines both VC and PC sub-engines. We
+ * allow to configure the tables through both, for the rare cases
+ * where a table only really needs to be configured for one of
+ * them (e.g. the NVG table for the presenter). It assumes that
+ * firmware passes the same address to the VC and PC when tables
+ * are defined for both, which seems acceptable.
*/
case PC_VSD_TABLE_ADDR:
+ break;
case PC_VSD_TABLE_DATA:
+ pnv_xive2_pc_vst_set_data(xive, val);
break;
case PC_NXC_PROC_CONFIG:
--
2.43.0
^ permalink raw reply related [flat|nested] 23+ messages in thread* Re: [PATCH v2 5/9] pnv/xive2: Configure Virtualization Structure Tables through the PC
2024-07-15 18:33 ` [PATCH v2 5/9] pnv/xive2: Configure Virtualization Structure Tables through the PC Michael Kowal
@ 2024-07-16 7:06 ` Cédric Le Goater
0 siblings, 0 replies; 23+ messages in thread
From: Cédric Le Goater @ 2024-07-16 7:06 UTC (permalink / raw)
To: Michael Kowal, qemu-devel; +Cc: qemu-ppc, fbarrat, npiggin, milesg
On 7/15/24 20:33, Michael Kowal wrote:
> From: Frederic Barrat <fbarrat@linux.ibm.com>
>
> Both the virtualization layer (VC) and presentation layer (PC) need to
> be configured to access the VSTs. Since the information is redundant,
> the xive model combines both into one set of tables and only the
> definitions going through the VC are kept. The definitions through the
> PC are ignored. That works well as long as firmware calls the VC for
> all the tables.
>
> For the NVG and NVC tables, it can make sense to only configure them
> with the PC, since they are only used by the presenter. So this patch
> allows firmware to configure the VST tables through the PC as well.
> The definitions are still shared, since the VST tables can be set
> through both the VC and/or PC, they are dynamically re-mapped in
> memory by first deleting the memory subregion.
>
> Signed-off-by: Frederic Barrat <fbarrat@linux.ibm.com>
> Signed-off-by: Michael Kowal <kowal@linux.vnet.ibm.com>
Reviewed-by: Cédric Le Goater <clg@redhat.com>
Thanks,
C.
> ---
> hw/intc/pnv_xive2.c | 47 ++++++++++++++++++++++++++++++++++++---------
> 1 file changed, 38 insertions(+), 9 deletions(-)
>
> diff --git a/hw/intc/pnv_xive2.c b/hw/intc/pnv_xive2.c
> index c72c66dd6a..9e45161869 100644
> --- a/hw/intc/pnv_xive2.c
> +++ b/hw/intc/pnv_xive2.c
> @@ -759,6 +759,9 @@ static void pnv_xive2_vst_set_exclusive(PnvXive2 *xive, uint8_t type,
> * entries provisioned by FW (such as skiboot) and resize the
> * ESB window accordingly.
> */
> + if (memory_region_is_mapped(&xsrc->esb_mmio)) {
> + memory_region_del_subregion(&xive->esb_mmio, &xsrc->esb_mmio);
> + }
> if (!(VSD_INDIRECT & vsd)) {
> memory_region_set_size(&xsrc->esb_mmio, vst_tsize * SBE_PER_BYTE
> * (1ull << xsrc->esb_shift));
> @@ -774,6 +777,9 @@ static void pnv_xive2_vst_set_exclusive(PnvXive2 *xive, uint8_t type,
> /*
> * Backing store pages for the END.
> */
> + if (memory_region_is_mapped(&end_xsrc->esb_mmio)) {
> + memory_region_del_subregion(&xive->end_mmio, &end_xsrc->esb_mmio);
> + }
> if (!(VSD_INDIRECT & vsd)) {
> memory_region_set_size(&end_xsrc->esb_mmio, (vst_tsize / info->size)
> * (1ull << end_xsrc->esb_shift));
> @@ -798,13 +804,10 @@ static void pnv_xive2_vst_set_exclusive(PnvXive2 *xive, uint8_t type,
> * Both PC and VC sub-engines are configured as each use the Virtual
> * Structure Tables
> */
> -static void pnv_xive2_vst_set_data(PnvXive2 *xive, uint64_t vsd)
> +static void pnv_xive2_vst_set_data(PnvXive2 *xive, uint64_t vsd,
> + uint8_t type, uint8_t blk)
> {
> uint8_t mode = GETFIELD(VSD_MODE, vsd);
> - uint8_t type = GETFIELD(VC_VSD_TABLE_SELECT,
> - xive->vc_regs[VC_VSD_TABLE_ADDR >> 3]);
> - uint8_t blk = GETFIELD(VC_VSD_TABLE_ADDRESS,
> - xive->vc_regs[VC_VSD_TABLE_ADDR >> 3]);
> uint64_t vst_addr = vsd & VSD_ADDRESS_MASK;
>
> if (type > VST_ERQ) {
> @@ -839,6 +842,16 @@ static void pnv_xive2_vst_set_data(PnvXive2 *xive, uint64_t vsd)
> }
> }
>
> +static void pnv_xive2_vc_vst_set_data(PnvXive2 *xive, uint64_t vsd)
> +{
> + uint8_t type = GETFIELD(VC_VSD_TABLE_SELECT,
> + xive->vc_regs[VC_VSD_TABLE_ADDR >> 3]);
> + uint8_t blk = GETFIELD(VC_VSD_TABLE_ADDRESS,
> + xive->vc_regs[VC_VSD_TABLE_ADDR >> 3]);
> +
> + pnv_xive2_vst_set_data(xive, vsd, type, blk);
> +}
> +
> /*
> * MMIO handlers
> */
> @@ -1268,7 +1281,7 @@ static void pnv_xive2_ic_vc_write(void *opaque, hwaddr offset,
> case VC_VSD_TABLE_ADDR:
> break;
> case VC_VSD_TABLE_DATA:
> - pnv_xive2_vst_set_data(xive, val);
> + pnv_xive2_vc_vst_set_data(xive, val);
> break;
>
> /*
> @@ -1487,6 +1500,16 @@ static uint64_t pnv_xive2_ic_pc_read(void *opaque, hwaddr offset,
> return val;
> }
>
> +static void pnv_xive2_pc_vst_set_data(PnvXive2 *xive, uint64_t vsd)
> +{
> + uint8_t type = GETFIELD(PC_VSD_TABLE_SELECT,
> + xive->pc_regs[PC_VSD_TABLE_ADDR >> 3]);
> + uint8_t blk = GETFIELD(PC_VSD_TABLE_ADDRESS,
> + xive->pc_regs[PC_VSD_TABLE_ADDR >> 3]);
> +
> + pnv_xive2_vst_set_data(xive, vsd, type, blk);
> +}
> +
> static void pnv_xive2_ic_pc_write(void *opaque, hwaddr offset,
> uint64_t val, unsigned size)
> {
> @@ -1497,12 +1520,18 @@ static void pnv_xive2_ic_pc_write(void *opaque, hwaddr offset,
> switch (offset) {
>
> /*
> - * VSD table settings. Only taken into account in the VC
> - * sub-engine because the Xive2Router model combines both VC and PC
> - * sub-engines
> + * VSD table settings.
> + * The Xive2Router model combines both VC and PC sub-engines. We
> + * allow to configure the tables through both, for the rare cases
> + * where a table only really needs to be configured for one of
> + * them (e.g. the NVG table for the presenter). It assumes that
> + * firmware passes the same address to the VC and PC when tables
> + * are defined for both, which seems acceptable.
> */
> case PC_VSD_TABLE_ADDR:
> + break;
> case PC_VSD_TABLE_DATA:
> + pnv_xive2_pc_vst_set_data(xive, val);
> break;
>
> case PC_NXC_PROC_CONFIG:
^ permalink raw reply [flat|nested] 23+ messages in thread
* [PATCH v2 6/9] pnv/xive2: Enable VST NVG and NVC index compression
2024-07-15 18:33 [PATCH v2 0/9] XIVE changes for Cache Watch, VSTs, STT and info pic Michael Kowal
` (4 preceding siblings ...)
2024-07-15 18:33 ` [PATCH v2 5/9] pnv/xive2: Configure Virtualization Structure Tables through the PC Michael Kowal
@ 2024-07-15 18:33 ` Michael Kowal
2024-07-16 7:05 ` Cédric Le Goater
2024-07-15 18:33 ` [PATCH v2 7/9] pnv/xive2: Set Translation Table for the NVC port space Michael Kowal
` (2 subsequent siblings)
8 siblings, 1 reply; 23+ messages in thread
From: Michael Kowal @ 2024-07-15 18:33 UTC (permalink / raw)
To: qemu-devel; +Cc: qemu-ppc, clg, fbarrat, npiggin, milesg
From: Frederic Barrat <fbarrat@linux.ibm.com>
Enable NVG and NVC VST tables for index compression which indicates the number
of bits the address is shifted to the right for the table accesses.
The compression values are defined as:
0000 - No compression
0001 - 1 bit shift
0010 - 2 bit shift
....
1000 - 8 bit shift
1001-1111 - No compression
Signed-off-by: Frederic Barrat <fbarrat@linux.ibm.com>
Signed-off-by: Michael Kowal <kowal@linux.vnet.ibm.com>
---
hw/intc/pnv_xive2_regs.h | 2 ++
hw/intc/pnv_xive2.c | 20 ++++++++++++++++++++
2 files changed, 22 insertions(+)
diff --git a/hw/intc/pnv_xive2_regs.h b/hw/intc/pnv_xive2_regs.h
index ca05255d20..e8b87b3d2c 100644
--- a/hw/intc/pnv_xive2_regs.h
+++ b/hw/intc/pnv_xive2_regs.h
@@ -427,6 +427,8 @@
#define X_PC_NXC_PROC_CONFIG 0x28A
#define PC_NXC_PROC_CONFIG 0x450
#define PC_NXC_PROC_CONFIG_WATCH_ASSIGN PPC_BITMASK(0, 3)
+#define PC_NXC_PROC_CONFIG_NVG_TABLE_COMPRESS PPC_BITMASK(32, 35)
+#define PC_NXC_PROC_CONFIG_NVC_TABLE_COMPRESS PPC_BITMASK(36, 39)
/* NxC Cache Watch 0 Specification */
#define X_PC_NXC_WATCH0_SPEC 0x2A0
diff --git a/hw/intc/pnv_xive2.c b/hw/intc/pnv_xive2.c
index 9e45161869..b14fad1b2e 100644
--- a/hw/intc/pnv_xive2.c
+++ b/hw/intc/pnv_xive2.c
@@ -217,6 +217,20 @@ static uint64_t pnv_xive2_vst_addr_indirect(PnvXive2 *xive, uint32_t type,
return pnv_xive2_vst_addr_direct(xive, type, vsd, (idx % vst_per_page));
}
+static uint8_t pnv_xive2_nvc_table_compress_shift(PnvXive2 *xive)
+{
+ uint8_t shift = GETFIELD(PC_NXC_PROC_CONFIG_NVC_TABLE_COMPRESS,
+ xive->pc_regs[PC_NXC_PROC_CONFIG >> 3]);
+ return shift > 8 ? 0 : shift;
+}
+
+static uint8_t pnv_xive2_nvg_table_compress_shift(PnvXive2 *xive)
+{
+ uint8_t shift = GETFIELD(PC_NXC_PROC_CONFIG_NVG_TABLE_COMPRESS,
+ xive->pc_regs[PC_NXC_PROC_CONFIG >> 3]);
+ return shift > 8 ? 0 : shift;
+}
+
static uint64_t pnv_xive2_vst_addr(PnvXive2 *xive, uint32_t type, uint8_t blk,
uint32_t idx)
{
@@ -238,6 +252,12 @@ static uint64_t pnv_xive2_vst_addr(PnvXive2 *xive, uint32_t type, uint8_t blk,
return xive ? pnv_xive2_vst_addr(xive, type, blk, idx) : 0;
}
+ if (type == VST_NVG) {
+ idx >>= pnv_xive2_nvg_table_compress_shift(xive);
+ } else if (type == VST_NVC) {
+ idx >>= pnv_xive2_nvc_table_compress_shift(xive);
+ }
+
if (VSD_INDIRECT & vsd) {
return pnv_xive2_vst_addr_indirect(xive, type, vsd, idx);
}
--
2.43.0
^ permalink raw reply related [flat|nested] 23+ messages in thread* Re: [PATCH v2 6/9] pnv/xive2: Enable VST NVG and NVC index compression
2024-07-15 18:33 ` [PATCH v2 6/9] pnv/xive2: Enable VST NVG and NVC index compression Michael Kowal
@ 2024-07-16 7:05 ` Cédric Le Goater
0 siblings, 0 replies; 23+ messages in thread
From: Cédric Le Goater @ 2024-07-16 7:05 UTC (permalink / raw)
To: Michael Kowal, qemu-devel; +Cc: qemu-ppc, fbarrat, npiggin, milesg
On 7/15/24 20:33, Michael Kowal wrote:
> From: Frederic Barrat <fbarrat@linux.ibm.com>
>
> Enable NVG and NVC VST tables for index compression which indicates the number
> of bits the address is shifted to the right for the table accesses.
> The compression values are defined as:
> 0000 - No compression
> 0001 - 1 bit shift
> 0010 - 2 bit shift
> ....
> 1000 - 8 bit shift
> 1001-1111 - No compression
>
> Signed-off-by: Frederic Barrat <fbarrat@linux.ibm.com>
> Signed-off-by: Michael Kowal <kowal@linux.vnet.ibm.com>
Reviewed-by: Cédric Le Goater <clg@redhat.com>
Thanks,
C.
> ---
> hw/intc/pnv_xive2_regs.h | 2 ++
> hw/intc/pnv_xive2.c | 20 ++++++++++++++++++++
> 2 files changed, 22 insertions(+)
>
> diff --git a/hw/intc/pnv_xive2_regs.h b/hw/intc/pnv_xive2_regs.h
> index ca05255d20..e8b87b3d2c 100644
> --- a/hw/intc/pnv_xive2_regs.h
> +++ b/hw/intc/pnv_xive2_regs.h
> @@ -427,6 +427,8 @@
> #define X_PC_NXC_PROC_CONFIG 0x28A
> #define PC_NXC_PROC_CONFIG 0x450
> #define PC_NXC_PROC_CONFIG_WATCH_ASSIGN PPC_BITMASK(0, 3)
> +#define PC_NXC_PROC_CONFIG_NVG_TABLE_COMPRESS PPC_BITMASK(32, 35)
> +#define PC_NXC_PROC_CONFIG_NVC_TABLE_COMPRESS PPC_BITMASK(36, 39)
>
> /* NxC Cache Watch 0 Specification */
> #define X_PC_NXC_WATCH0_SPEC 0x2A0
> diff --git a/hw/intc/pnv_xive2.c b/hw/intc/pnv_xive2.c
> index 9e45161869..b14fad1b2e 100644
> --- a/hw/intc/pnv_xive2.c
> +++ b/hw/intc/pnv_xive2.c
> @@ -217,6 +217,20 @@ static uint64_t pnv_xive2_vst_addr_indirect(PnvXive2 *xive, uint32_t type,
> return pnv_xive2_vst_addr_direct(xive, type, vsd, (idx % vst_per_page));
> }
>
> +static uint8_t pnv_xive2_nvc_table_compress_shift(PnvXive2 *xive)
> +{
> + uint8_t shift = GETFIELD(PC_NXC_PROC_CONFIG_NVC_TABLE_COMPRESS,
> + xive->pc_regs[PC_NXC_PROC_CONFIG >> 3]);
> + return shift > 8 ? 0 : shift;
> +}
> +
> +static uint8_t pnv_xive2_nvg_table_compress_shift(PnvXive2 *xive)
> +{
> + uint8_t shift = GETFIELD(PC_NXC_PROC_CONFIG_NVG_TABLE_COMPRESS,
> + xive->pc_regs[PC_NXC_PROC_CONFIG >> 3]);
> + return shift > 8 ? 0 : shift;
> +}
> +
> static uint64_t pnv_xive2_vst_addr(PnvXive2 *xive, uint32_t type, uint8_t blk,
> uint32_t idx)
> {
> @@ -238,6 +252,12 @@ static uint64_t pnv_xive2_vst_addr(PnvXive2 *xive, uint32_t type, uint8_t blk,
> return xive ? pnv_xive2_vst_addr(xive, type, blk, idx) : 0;
> }
>
> + if (type == VST_NVG) {
> + idx >>= pnv_xive2_nvg_table_compress_shift(xive);
> + } else if (type == VST_NVC) {
> + idx >>= pnv_xive2_nvc_table_compress_shift(xive);
> + }
> +
> if (VSD_INDIRECT & vsd) {
> return pnv_xive2_vst_addr_indirect(xive, type, vsd, idx);
> }
^ permalink raw reply [flat|nested] 23+ messages in thread
* [PATCH v2 7/9] pnv/xive2: Set Translation Table for the NVC port space
2024-07-15 18:33 [PATCH v2 0/9] XIVE changes for Cache Watch, VSTs, STT and info pic Michael Kowal
` (5 preceding siblings ...)
2024-07-15 18:33 ` [PATCH v2 6/9] pnv/xive2: Enable VST NVG and NVC index compression Michael Kowal
@ 2024-07-15 18:33 ` Michael Kowal
2024-07-16 7:07 ` Cédric Le Goater
2024-07-15 18:33 ` [PATCH v2 8/9] pnv/xive2: Fail VST entry address computation if table has no VSD Michael Kowal
2024-07-15 18:33 ` [PATCH v2 9/9] pnv/xive2: Move xive2_nvp_pic_print_info() to xive2.c Michael Kowal
8 siblings, 1 reply; 23+ messages in thread
From: Michael Kowal @ 2024-07-15 18:33 UTC (permalink / raw)
To: qemu-devel; +Cc: qemu-ppc, clg, fbarrat, npiggin, milesg
From: Frederic Barrat <fbarrat@linux.ibm.com>
Set Translation Table for the NVC port space is missing. The xive model
doesn't take into account the remapping of IO operations via the Set
Translation Table but firmware is allowed to define it for the Notify
Virtual Crowd (NVC), like it's already done for the other VST tables.
Signed-off-by: Frederic Barrat <fbarrat@linux.ibm.com>
Signed-off-by: Michael Kowal <kowal@linux.vnet.ibm.com>
---
hw/intc/pnv_xive2.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/hw/intc/pnv_xive2.c b/hw/intc/pnv_xive2.c
index b14fad1b2e..24ecab46a1 100644
--- a/hw/intc/pnv_xive2.c
+++ b/hw/intc/pnv_xive2.c
@@ -719,6 +719,7 @@ static int pnv_xive2_stt_set_data(PnvXive2 *xive, uint64_t val)
case CQ_TAR_NVPG:
case CQ_TAR_ESB:
case CQ_TAR_END:
+ case CQ_TAR_NVC:
xive->tables[tsel][entry] = val;
break;
default:
--
2.43.0
^ permalink raw reply related [flat|nested] 23+ messages in thread* Re: [PATCH v2 7/9] pnv/xive2: Set Translation Table for the NVC port space
2024-07-15 18:33 ` [PATCH v2 7/9] pnv/xive2: Set Translation Table for the NVC port space Michael Kowal
@ 2024-07-16 7:07 ` Cédric Le Goater
0 siblings, 0 replies; 23+ messages in thread
From: Cédric Le Goater @ 2024-07-16 7:07 UTC (permalink / raw)
To: Michael Kowal, qemu-devel; +Cc: qemu-ppc, fbarrat, npiggin, milesg
On 7/15/24 20:33, Michael Kowal wrote:
> From: Frederic Barrat <fbarrat@linux.ibm.com>
>
> Set Translation Table for the NVC port space is missing. The xive model
> doesn't take into account the remapping of IO operations via the Set
> Translation Table but firmware is allowed to define it for the Notify
> Virtual Crowd (NVC), like it's already done for the other VST tables.
>
> Signed-off-by: Frederic Barrat <fbarrat@linux.ibm.com>
> Signed-off-by: Michael Kowal <kowal@linux.vnet.ibm.com>
Reviewed-by: Cédric Le Goater <clg@redhat.com>
Thanks,
C.
> ---
> hw/intc/pnv_xive2.c | 1 +
> 1 file changed, 1 insertion(+)
>
> diff --git a/hw/intc/pnv_xive2.c b/hw/intc/pnv_xive2.c
> index b14fad1b2e..24ecab46a1 100644
> --- a/hw/intc/pnv_xive2.c
> +++ b/hw/intc/pnv_xive2.c
> @@ -719,6 +719,7 @@ static int pnv_xive2_stt_set_data(PnvXive2 *xive, uint64_t val)
> case CQ_TAR_NVPG:
> case CQ_TAR_ESB:
> case CQ_TAR_END:
> + case CQ_TAR_NVC:
> xive->tables[tsel][entry] = val;
> break;
> default:
^ permalink raw reply [flat|nested] 23+ messages in thread
* [PATCH v2 8/9] pnv/xive2: Fail VST entry address computation if table has no VSD
2024-07-15 18:33 [PATCH v2 0/9] XIVE changes for Cache Watch, VSTs, STT and info pic Michael Kowal
` (6 preceding siblings ...)
2024-07-15 18:33 ` [PATCH v2 7/9] pnv/xive2: Set Translation Table for the NVC port space Michael Kowal
@ 2024-07-15 18:33 ` Michael Kowal
2024-07-16 7:05 ` Cédric Le Goater
2024-07-15 18:33 ` [PATCH v2 9/9] pnv/xive2: Move xive2_nvp_pic_print_info() to xive2.c Michael Kowal
8 siblings, 1 reply; 23+ messages in thread
From: Michael Kowal @ 2024-07-15 18:33 UTC (permalink / raw)
To: qemu-devel; +Cc: qemu-ppc, clg, fbarrat, npiggin, milesg
From: Frederic Barrat <fbarrat@linux.ibm.com>
Fail VST entry address computation if firmware doesn't define a descriptor
for one of the Virtualization Structure Tables (VST), there's no point in
trying to compute the address of its entry. Abort the operation and log
an error.
Signed-off-by: Frederic Barrat <fbarrat@linux.ibm.com>
Signed-off-by: Michael Kowal <kowal@linux.vnet.ibm.com>
---
hw/intc/pnv_xive2.c | 5 +++++
1 file changed, 5 insertions(+)
diff --git a/hw/intc/pnv_xive2.c b/hw/intc/pnv_xive2.c
index 24ecab46a1..ad0d6f45e4 100644
--- a/hw/intc/pnv_xive2.c
+++ b/hw/intc/pnv_xive2.c
@@ -244,6 +244,11 @@ static uint64_t pnv_xive2_vst_addr(PnvXive2 *xive, uint32_t type, uint8_t blk,
}
vsd = xive->vsds[type][blk];
+ if (vsd == 0) {
+ xive2_error(xive, "VST: vsd == 0 block id %d for VST %s %d !?",
+ blk, info->name, idx);
+ return 0;
+ }
/* Remote VST access */
if (GETFIELD(VSD_MODE, vsd) == VSD_MODE_FORWARD) {
--
2.43.0
^ permalink raw reply related [flat|nested] 23+ messages in thread* Re: [PATCH v2 8/9] pnv/xive2: Fail VST entry address computation if table has no VSD
2024-07-15 18:33 ` [PATCH v2 8/9] pnv/xive2: Fail VST entry address computation if table has no VSD Michael Kowal
@ 2024-07-16 7:05 ` Cédric Le Goater
0 siblings, 0 replies; 23+ messages in thread
From: Cédric Le Goater @ 2024-07-16 7:05 UTC (permalink / raw)
To: Michael Kowal, qemu-devel; +Cc: qemu-ppc, fbarrat, npiggin, milesg
On 7/15/24 20:33, Michael Kowal wrote:
> From: Frederic Barrat <fbarrat@linux.ibm.com>
>
> Fail VST entry address computation if firmware doesn't define a descriptor
> for one of the Virtualization Structure Tables (VST), there's no point in
> trying to compute the address of its entry. Abort the operation and log
> an error.
>
> Signed-off-by: Frederic Barrat <fbarrat@linux.ibm.com>
> Signed-off-by: Michael Kowal <kowal@linux.vnet.ibm.com>
Reviewed-by: Cédric Le Goater <clg@redhat.com>
Thanks,
C.
> ---
> hw/intc/pnv_xive2.c | 5 +++++
> 1 file changed, 5 insertions(+)
>
> diff --git a/hw/intc/pnv_xive2.c b/hw/intc/pnv_xive2.c
> index 24ecab46a1..ad0d6f45e4 100644
> --- a/hw/intc/pnv_xive2.c
> +++ b/hw/intc/pnv_xive2.c
> @@ -244,6 +244,11 @@ static uint64_t pnv_xive2_vst_addr(PnvXive2 *xive, uint32_t type, uint8_t blk,
> }
>
> vsd = xive->vsds[type][blk];
> + if (vsd == 0) {
> + xive2_error(xive, "VST: vsd == 0 block id %d for VST %s %d !?",
> + blk, info->name, idx);
> + return 0;
> + }
>
> /* Remote VST access */
> if (GETFIELD(VSD_MODE, vsd) == VSD_MODE_FORWARD) {
^ permalink raw reply [flat|nested] 23+ messages in thread
* [PATCH v2 9/9] pnv/xive2: Move xive2_nvp_pic_print_info() to xive2.c
2024-07-15 18:33 [PATCH v2 0/9] XIVE changes for Cache Watch, VSTs, STT and info pic Michael Kowal
` (7 preceding siblings ...)
2024-07-15 18:33 ` [PATCH v2 8/9] pnv/xive2: Fail VST entry address computation if table has no VSD Michael Kowal
@ 2024-07-15 18:33 ` Michael Kowal
2024-07-16 7:15 ` Cédric Le Goater
8 siblings, 1 reply; 23+ messages in thread
From: Michael Kowal @ 2024-07-15 18:33 UTC (permalink / raw)
To: qemu-devel; +Cc: qemu-ppc, clg, fbarrat, npiggin, milesg
From: Frederic Barrat <fbarrat@linux.ibm.com>
Moving xive2_nvp_pic_print_info() align with the other "pic_print_info"
functions and allows us to call functions internal to xive2.c.
In XIVE Gen 2 there were some minor changes to the TIMA header that were
updated when printed.
Additional END state 'info pic' information as added. The 'ignore',
'crowd' and 'precluded escalation control' bits of an Event Notification
Descriptor are all used when delivering an interrupt targeting a VP-group
or crowd.
Signed-off-by: Frederic Barrat <fbarrat@linux.ibm.com>
Signed-off-by: Michael Kowal <kowal@linux.vnet.ibm.com>
---
include/hw/ppc/xive2_regs.h | 9 +++++++++
hw/intc/pnv_xive2.c | 27 ---------------------------
hw/intc/xive.c | 12 +++++++++---
hw/intc/xive2.c | 33 +++++++++++++++++++++++++++++++--
4 files changed, 49 insertions(+), 32 deletions(-)
diff --git a/include/hw/ppc/xive2_regs.h b/include/hw/ppc/xive2_regs.h
index 4e5e17cd89..4349d009d0 100644
--- a/include/hw/ppc/xive2_regs.h
+++ b/include/hw/ppc/xive2_regs.h
@@ -97,6 +97,7 @@ typedef struct Xive2End {
uint32_t w6;
#define END2_W6_FORMAT_BIT PPC_BIT32(0)
#define END2_W6_IGNORE PPC_BIT32(1)
+#define END2_W6_CROWD PPC_BIT32(2)
#define END2_W6_VP_BLOCK PPC_BITMASK32(4, 7)
#define END2_W6_VP_OFFSET PPC_BITMASK32(8, 31)
#define END2_W6_VP_OFFSET_GEN1 PPC_BITMASK32(13, 31)
@@ -111,6 +112,8 @@ typedef struct Xive2End {
#define xive2_end_is_notify(end) \
(be32_to_cpu((end)->w0) & END2_W0_UCOND_NOTIFY)
#define xive2_end_is_backlog(end) (be32_to_cpu((end)->w0) & END2_W0_BACKLOG)
+#define xive2_end_is_precluded_escalation(end) \
+ (be32_to_cpu((end)->w0) & END2_W0_PRECL_ESC_CTL)
#define xive2_end_is_escalate(end) \
(be32_to_cpu((end)->w0) & END2_W0_ESCALATE_CTL)
#define xive2_end_is_uncond_escalation(end) \
@@ -123,6 +126,10 @@ typedef struct Xive2End {
(be32_to_cpu((end)->w0) & END2_W0_FIRMWARE1)
#define xive2_end_is_firmware2(end) \
(be32_to_cpu((end)->w0) & END2_W0_FIRMWARE2)
+#define xive2_end_is_ignore(end) \
+ (be32_to_cpu((end)->w6) & END2_W6_IGNORE)
+#define xive2_end_is_crowd(end) \
+ (be32_to_cpu((end)->w6) & END2_W6_CROWD)
static inline uint64_t xive2_end_qaddr(Xive2End *end)
{
@@ -194,6 +201,8 @@ static inline uint32_t xive2_nvp_blk(uint32_t cam_line)
return (cam_line >> XIVE2_NVP_SHIFT) & 0xf;
}
+void xive2_nvp_pic_print_info(Xive2Nvp *nvp, uint32_t nvp_idx, GString *buf);
+
/*
* Notification Virtual Group or Crowd (NVG/NVC)
*/
diff --git a/hw/intc/pnv_xive2.c b/hw/intc/pnv_xive2.c
index ad0d6f45e4..7846f2ef33 100644
--- a/hw/intc/pnv_xive2.c
+++ b/hw/intc/pnv_xive2.c
@@ -2433,33 +2433,6 @@ static void pnv_xive2_register_types(void)
type_init(pnv_xive2_register_types)
-static void xive2_nvp_pic_print_info(Xive2Nvp *nvp, uint32_t nvp_idx,
- GString *buf)
-{
- uint8_t eq_blk = xive_get_field32(NVP2_W5_VP_END_BLOCK, nvp->w5);
- uint32_t eq_idx = xive_get_field32(NVP2_W5_VP_END_INDEX, nvp->w5);
-
- if (!xive2_nvp_is_valid(nvp)) {
- return;
- }
-
- g_string_append_printf(buf, " %08x end:%02x/%04x IPB:%02x",
- nvp_idx, eq_blk, eq_idx,
- xive_get_field32(NVP2_W2_IPB, nvp->w2));
- /*
- * When the NVP is HW controlled, more fields are updated
- */
- if (xive2_nvp_is_hw(nvp)) {
- g_string_append_printf(buf, " CPPR:%02x",
- xive_get_field32(NVP2_W2_CPPR, nvp->w2));
- if (xive2_nvp_is_co(nvp)) {
- g_string_append_printf(buf, " CO:%04x",
- xive_get_field32(NVP2_W1_CO_THRID, nvp->w1));
- }
- }
- g_string_append_c(buf, '\n');
-}
-
/*
* If the table is direct, we can compute the number of PQ entries
* provisioned by FW.
diff --git a/hw/intc/xive.c b/hw/intc/xive.c
index 70f11f993b..5a02dd8e02 100644
--- a/hw/intc/xive.c
+++ b/hw/intc/xive.c
@@ -692,9 +692,15 @@ void xive_tctx_pic_print_info(XiveTCTX *tctx, GString *buf)
}
}
- g_string_append_printf(buf, "CPU[%04x]: "
- "QW NSR CPPR IPB LSMFB ACK# INC AGE PIPR W2\n",
- cpu_index);
+ if (xive_presenter_get_config(tctx->xptr) & XIVE_PRESENTER_GEN1_TIMA_OS) {
+ g_string_append_printf(buf, "CPU[%04x]: "
+ "QW NSR CPPR IPB LSMFB ACK# INC AGE PIPR"
+ " W2\n", cpu_index);
+ } else {
+ g_string_append_printf(buf, "CPU[%04x]: "
+ "QW NSR CPPR IPB LSMFB - LGS T PIPR"
+ " W2\n", cpu_index);
+ }
for (i = 0; i < XIVE_TM_RING_COUNT; i++) {
char *s = xive_tctx_ring_print(&tctx->regs[i * XIVE_TM_RING_SIZE]);
diff --git a/hw/intc/xive2.c b/hw/intc/xive2.c
index 3e7238c663..1f150685bf 100644
--- a/hw/intc/xive2.c
+++ b/hw/intc/xive2.c
@@ -89,7 +89,7 @@ void xive2_end_pic_print_info(Xive2End *end, uint32_t end_idx, GString *buf)
pq = xive_get_field32(END2_W1_ESn, end->w1);
g_string_append_printf(buf,
- " %08x %c%c %c%c%c%c%c%c%c%c%c%c "
+ " %08x %c%c %c%c%c%c%c%c%c%c%c%c%c %c%c "
"prio:%d nvp:%02x/%04x",
end_idx,
pq & XIVE_ESB_VAL_P ? 'P' : '-',
@@ -98,12 +98,15 @@ void xive2_end_pic_print_info(Xive2End *end, uint32_t end_idx, GString *buf)
xive2_end_is_enqueue(end) ? 'q' : '-',
xive2_end_is_notify(end) ? 'n' : '-',
xive2_end_is_backlog(end) ? 'b' : '-',
+ xive2_end_is_precluded_escalation(end) ? 'p' : '-',
xive2_end_is_escalate(end) ? 'e' : '-',
xive2_end_is_escalate_end(end) ? 'N' : '-',
xive2_end_is_uncond_escalation(end) ? 'u' : '-',
xive2_end_is_silent_escalation(end) ? 's' : '-',
xive2_end_is_firmware1(end) ? 'f' : '-',
xive2_end_is_firmware2(end) ? 'F' : '-',
+ xive2_end_is_ignore(end) ? 'i' : '-',
+ xive2_end_is_crowd(end) ? 'c' : '-',
priority, nvp_blk, nvp_idx);
if (qaddr_base) {
@@ -137,6 +140,32 @@ void xive2_end_eas_pic_print_info(Xive2End *end, uint32_t end_idx,
(uint32_t) xive_get_field64(EAS2_END_DATA, eas->w));
}
+void xive2_nvp_pic_print_info(Xive2Nvp *nvp, uint32_t nvp_idx, GString *buf)
+{
+ uint8_t eq_blk = xive_get_field32(NVP2_W5_VP_END_BLOCK, nvp->w5);
+ uint32_t eq_idx = xive_get_field32(NVP2_W5_VP_END_INDEX, nvp->w5);
+
+ if (!xive2_nvp_is_valid(nvp)) {
+ return;
+ }
+
+ g_string_append_printf(buf, " %08x end:%02x/%04x IPB:%02x",
+ nvp_idx, eq_blk, eq_idx,
+ xive_get_field32(NVP2_W2_IPB, nvp->w2));
+ /*
+ * When the NVP is HW controlled, more fields are updated
+ */
+ if (xive2_nvp_is_hw(nvp)) {
+ g_string_append_printf(buf, " CPPR:%02x",
+ xive_get_field32(NVP2_W2_CPPR, nvp->w2));
+ if (xive2_nvp_is_co(nvp)) {
+ g_string_append_printf(buf, " CO:%04x",
+ xive_get_field32(NVP2_W1_CO_THRID, nvp->w1));
+ }
+ }
+ g_string_append_c(buf, '\n');
+}
+
static void xive2_end_enqueue(Xive2End *end, uint32_t data)
{
uint64_t qaddr_base = xive2_end_qaddr(end);
@@ -650,7 +679,7 @@ static void xive2_router_end_notify(Xive2Router *xrtr, uint8_t end_blk,
}
found = xive_presenter_notify(xrtr->xfb, format, nvp_blk, nvp_idx,
- xive_get_field32(END2_W6_IGNORE, end.w7),
+ xive2_end_is_ignore(&end),
priority,
xive_get_field32(END2_W7_F1_LOG_SERVER_ID, end.w7));
--
2.43.0
^ permalink raw reply related [flat|nested] 23+ messages in thread* Re: [PATCH v2 9/9] pnv/xive2: Move xive2_nvp_pic_print_info() to xive2.c
2024-07-15 18:33 ` [PATCH v2 9/9] pnv/xive2: Move xive2_nvp_pic_print_info() to xive2.c Michael Kowal
@ 2024-07-16 7:15 ` Cédric Le Goater
0 siblings, 0 replies; 23+ messages in thread
From: Cédric Le Goater @ 2024-07-16 7:15 UTC (permalink / raw)
To: Michael Kowal, qemu-devel; +Cc: qemu-ppc, fbarrat, npiggin, milesg
On 7/15/24 20:33, Michael Kowal wrote:
> From: Frederic Barrat <fbarrat@linux.ibm.com>
>
> Moving xive2_nvp_pic_print_info() align with the other "pic_print_info"
> functions and allows us to call functions internal to xive2.c.
>
> In XIVE Gen 2 there were some minor changes to the TIMA header that were
> updated when printed.
>
> Additional END state 'info pic' information as added. The 'ignore',
> 'crowd' and 'precluded escalation control' bits of an Event Notification
> Descriptor are all used when delivering an interrupt targeting a VP-group
> or crowd.
>
> Signed-off-by: Frederic Barrat <fbarrat@linux.ibm.com>
> Signed-off-by: Michael Kowal <kowal@linux.vnet.ibm.com>
There are multiple unrelated changes in this patch.
> ---
> include/hw/ppc/xive2_regs.h | 9 +++++++++
> hw/intc/pnv_xive2.c | 27 ---------------------------
> hw/intc/xive.c | 12 +++++++++---
> hw/intc/xive2.c | 33 +++++++++++++++++++++++++++++++--
> 4 files changed, 49 insertions(+), 32 deletions(-)
>
> diff --git a/include/hw/ppc/xive2_regs.h b/include/hw/ppc/xive2_regs.h
> index 4e5e17cd89..4349d009d0 100644
> --- a/include/hw/ppc/xive2_regs.h
> +++ b/include/hw/ppc/xive2_regs.h
> @@ -97,6 +97,7 @@ typedef struct Xive2End {
> uint32_t w6;
> #define END2_W6_FORMAT_BIT PPC_BIT32(0)
> #define END2_W6_IGNORE PPC_BIT32(1)
> +#define END2_W6_CROWD PPC_BIT32(2)
> #define END2_W6_VP_BLOCK PPC_BITMASK32(4, 7)
> #define END2_W6_VP_OFFSET PPC_BITMASK32(8, 31)
> #define END2_W6_VP_OFFSET_GEN1 PPC_BITMASK32(13, 31)
> @@ -111,6 +112,8 @@ typedef struct Xive2End {
> #define xive2_end_is_notify(end) \
> (be32_to_cpu((end)->w0) & END2_W0_UCOND_NOTIFY)
> #define xive2_end_is_backlog(end) (be32_to_cpu((end)->w0) & END2_W0_BACKLOG)
> +#define xive2_end_is_precluded_escalation(end) \
> + (be32_to_cpu((end)->w0) & END2_W0_PRECL_ESC_CTL)
> #define xive2_end_is_escalate(end) \
> (be32_to_cpu((end)->w0) & END2_W0_ESCALATE_CTL)
> #define xive2_end_is_uncond_escalation(end) \
> @@ -123,6 +126,10 @@ typedef struct Xive2End {
> (be32_to_cpu((end)->w0) & END2_W0_FIRMWARE1)
> #define xive2_end_is_firmware2(end) \
> (be32_to_cpu((end)->w0) & END2_W0_FIRMWARE2)
> +#define xive2_end_is_ignore(end) \
> + (be32_to_cpu((end)->w6) & END2_W6_IGNORE)
> +#define xive2_end_is_crowd(end) \
> + (be32_to_cpu((end)->w6) & END2_W6_CROWD)
This is a first patch, plus the xive2_end_pic_print_info and
xive2_router_end_notify changes
> static inline uint64_t xive2_end_qaddr(Xive2End *end)
> {
> @@ -194,6 +201,8 @@ static inline uint32_t xive2_nvp_blk(uint32_t cam_line)
> return (cam_line >> XIVE2_NVP_SHIFT) & 0xf;
> }
>
> +void xive2_nvp_pic_print_info(Xive2Nvp *nvp, uint32_t nvp_idx, GString *buf);
> +
> /*
> * Notification Virtual Group or Crowd (NVG/NVC)
> */
> diff --git a/hw/intc/pnv_xive2.c b/hw/intc/pnv_xive2.c
> index ad0d6f45e4..7846f2ef33 100644
> --- a/hw/intc/pnv_xive2.c
> +++ b/hw/intc/pnv_xive2.c
> @@ -2433,33 +2433,6 @@ static void pnv_xive2_register_types(void)
>
> type_init(pnv_xive2_register_types)
>
> -static void xive2_nvp_pic_print_info(Xive2Nvp *nvp, uint32_t nvp_idx,
> - GString *buf)
> -{
> - uint8_t eq_blk = xive_get_field32(NVP2_W5_VP_END_BLOCK, nvp->w5);
> - uint32_t eq_idx = xive_get_field32(NVP2_W5_VP_END_INDEX, nvp->w5);
> -
> - if (!xive2_nvp_is_valid(nvp)) {
> - return;
> - }
> -
> - g_string_append_printf(buf, " %08x end:%02x/%04x IPB:%02x",
> - nvp_idx, eq_blk, eq_idx,
> - xive_get_field32(NVP2_W2_IPB, nvp->w2));
> - /*
> - * When the NVP is HW controlled, more fields are updated
> - */
> - if (xive2_nvp_is_hw(nvp)) {
> - g_string_append_printf(buf, " CPPR:%02x",
> - xive_get_field32(NVP2_W2_CPPR, nvp->w2));
> - if (xive2_nvp_is_co(nvp)) {
> - g_string_append_printf(buf, " CO:%04x",
> - xive_get_field32(NVP2_W1_CO_THRID, nvp->w1));
> - }
> - }
> - g_string_append_c(buf, '\n');
> -}
This is a second patch but I don't see where else xive2_nvp_pic_print_info()
is used ?
> /*
> * If the table is direct, we can compute the number of PQ entries
> * provisioned by FW.
> diff --git a/hw/intc/xive.c b/hw/intc/xive.c
> index 70f11f993b..5a02dd8e02 100644
> --- a/hw/intc/xive.c
> +++ b/hw/intc/xive.c
> @@ -692,9 +692,15 @@ void xive_tctx_pic_print_info(XiveTCTX *tctx, GString *buf)
> }
> }
>
> - g_string_append_printf(buf, "CPU[%04x]: "
> - "QW NSR CPPR IPB LSMFB ACK# INC AGE PIPR W2\n",
> - cpu_index);
> + if (xive_presenter_get_config(tctx->xptr) & XIVE_PRESENTER_GEN1_TIMA_OS) {
> + g_string_append_printf(buf, "CPU[%04x]: "
> + "QW NSR CPPR IPB LSMFB ACK# INC AGE PIPR"
> + " W2\n", cpu_index);
> + } else {
> + g_string_append_printf(buf, "CPU[%04x]: "
> + "QW NSR CPPR IPB LSMFB - LGS T PIPR"
> + " W2\n", cpu_index);
> + }
and this is patch 3.
Thanks,
C.
> for (i = 0; i < XIVE_TM_RING_COUNT; i++) {
> char *s = xive_tctx_ring_print(&tctx->regs[i * XIVE_TM_RING_SIZE]);
> diff --git a/hw/intc/xive2.c b/hw/intc/xive2.c
> index 3e7238c663..1f150685bf 100644
> --- a/hw/intc/xive2.c
> +++ b/hw/intc/xive2.c
> @@ -89,7 +89,7 @@ void xive2_end_pic_print_info(Xive2End *end, uint32_t end_idx, GString *buf)
> pq = xive_get_field32(END2_W1_ESn, end->w1);
>
> g_string_append_printf(buf,
> - " %08x %c%c %c%c%c%c%c%c%c%c%c%c "
> + " %08x %c%c %c%c%c%c%c%c%c%c%c%c%c %c%c "
> "prio:%d nvp:%02x/%04x",
> end_idx,
> pq & XIVE_ESB_VAL_P ? 'P' : '-',
> @@ -98,12 +98,15 @@ void xive2_end_pic_print_info(Xive2End *end, uint32_t end_idx, GString *buf)
> xive2_end_is_enqueue(end) ? 'q' : '-',
> xive2_end_is_notify(end) ? 'n' : '-',
> xive2_end_is_backlog(end) ? 'b' : '-',
> + xive2_end_is_precluded_escalation(end) ? 'p' : '-',
> xive2_end_is_escalate(end) ? 'e' : '-',
> xive2_end_is_escalate_end(end) ? 'N' : '-',
> xive2_end_is_uncond_escalation(end) ? 'u' : '-',
> xive2_end_is_silent_escalation(end) ? 's' : '-',
> xive2_end_is_firmware1(end) ? 'f' : '-',
> xive2_end_is_firmware2(end) ? 'F' : '-',
> + xive2_end_is_ignore(end) ? 'i' : '-',
> + xive2_end_is_crowd(end) ? 'c' : '-',
> priority, nvp_blk, nvp_idx);
>
> if (qaddr_base) {
> @@ -137,6 +140,32 @@ void xive2_end_eas_pic_print_info(Xive2End *end, uint32_t end_idx,
> (uint32_t) xive_get_field64(EAS2_END_DATA, eas->w));
> }
>
> +void xive2_nvp_pic_print_info(Xive2Nvp *nvp, uint32_t nvp_idx, GString *buf)
> +{
> + uint8_t eq_blk = xive_get_field32(NVP2_W5_VP_END_BLOCK, nvp->w5);
> + uint32_t eq_idx = xive_get_field32(NVP2_W5_VP_END_INDEX, nvp->w5);
> +
> + if (!xive2_nvp_is_valid(nvp)) {
> + return;
> + }
> +
> + g_string_append_printf(buf, " %08x end:%02x/%04x IPB:%02x",
> + nvp_idx, eq_blk, eq_idx,
> + xive_get_field32(NVP2_W2_IPB, nvp->w2));
> + /*
> + * When the NVP is HW controlled, more fields are updated
> + */
> + if (xive2_nvp_is_hw(nvp)) {
> + g_string_append_printf(buf, " CPPR:%02x",
> + xive_get_field32(NVP2_W2_CPPR, nvp->w2));
> + if (xive2_nvp_is_co(nvp)) {
> + g_string_append_printf(buf, " CO:%04x",
> + xive_get_field32(NVP2_W1_CO_THRID, nvp->w1));
> + }
> + }
> + g_string_append_c(buf, '\n');
> +}
> +
> static void xive2_end_enqueue(Xive2End *end, uint32_t data)
> {
> uint64_t qaddr_base = xive2_end_qaddr(end);
> @@ -650,7 +679,7 @@ static void xive2_router_end_notify(Xive2Router *xrtr, uint8_t end_blk,
> }
>
> found = xive_presenter_notify(xrtr->xfb, format, nvp_blk, nvp_idx,
> - xive_get_field32(END2_W6_IGNORE, end.w7),
> + xive2_end_is_ignore(&end),
> priority,
> xive_get_field32(END2_W7_F1_LOG_SERVER_ID, end.w7));
>
^ permalink raw reply [flat|nested] 23+ messages in thread