Skip to contentSkip to conversation

cxl/plugins: Hotness Monitoring Unit with 'real' data.

RFC PATCH In review Jonathan Cameron 3 participants 13 messages Jan 24, 2025 – May 20, 2025 Arm/Amlogic Meson SoC support

lore
Share & tools
lore
#0 lore raw

cover cxl/plugins: Hotness Monitoring Unit with 'real' data.

Hi All,

This is an RFC mainly to seek feedback on the approach used, particularly the aspect of how to get data from a TCG plugin into a device model. Two options that we have tried
1. Socket over which the plugin sends data to an external server (as seen here)
2. Register and manage a plugin from within a device model

The external server approach keeps things loosely coupled, but at the cost of separately maintaining that server, protocol definitions etc and some overhead.
The closely couple solution is neater, but I suspect might be controversial (hence I didn't start with that :)

The code here is at best a PoC to illustrate what we have in mind It's not nice code at all, feature gaps, bugs and all! So whilst review is always welcome I'm not requesting it for now.

Kernel support was posted a while back but was done against fake data (still supported here if you don't provide the port parameter to the type3 device)
https://lore.kernel.org/linux-cxl/[email protected]/
I'll post a minor update of that driver shortly to take into account a few specification clarifications but it should work with this without those.

Note there are some other patches on the tree I generated this from so this may not apply to upstream. Easiest is probably to test using gitlab.com/jic23/qemu cxl-2025-01-24

Thanks to Niyas for his suggestions on how to make all this work!

Background
----------

What is the Compute eXpress Link Hotness Monitoring unit and what is it for?
- In a tiered memory equipped server with the slow tier being attached via CXL the expectation is a given workload will benefit from putting data that is frequently fetched from memory in lower latency directly attached DRAM. Less frequently used data can be served from the CXL attached memory with no significant loss of performance. Any data that is hot enough to almost always be in cache doesn't matter as it is only fetch from memory occasionally.
- Working out which memory is best places where is hard to do and in some workloads a dynamic problem. As such we need something we can measure to provide some indication of what data is in the wrong place. There are existing techniques to do this (page faulting, various CPU tracing systems, access bit scanning etc) but they all have significant overheads.
- Monitoring accesses on the CXL device provides a path to getting good data without those overheads. These units are known as CXL Hotness Monitoring Units or CHMUs. Loosely speaking they count accesses to granuals of data (e.g. 4KiB pages). Exactly how they do that and where they sacrifice data accuracy is an implementation trade off.

Why do we need a model that gives real data?
- In general there is a need to develop software on top of these units to move data to the right place. Hard to evaluate that if we are making up the info on what is 'hot'.
- Need to allow for a bunch of 'impdef' solutions. Note that CHMU in this patch set is an oracle - it has enough counters to count every access. That's not realistic but it doesn't get me shouted at by our architecture teams for giving away any secrets. If we move forward with this, I'll probably implement a limited counter + full CAM solution (also unrealistic, but closer to real) I'd be very interested in contributions of other approaches (there are lots in the literature, under the term top-k)
- Resources will be constrained, so whilst a CHMU might in theory allow monitoring everything at once, that will come with a big accuracy cost. We need to design the algorithms that give us good data given those constraints.

So we need a solution to explore the design space and develop the software to take advantage of this hardware (there are various LSF/MM proposals on how to use this an other ways of tracking hotness).
https://lore.kernel.org/all/[email protected]/
https://lore.kernel.org/lkml/Z4XUoWlU-UgRik18@gourry-fedora-PF4VCD3F/T/

QEMU plugins give us a way to do this. In particular the existing Cache plugin can be easily modified to tell use what memory addresses missed at the last level of emulated cache. We can then filter those for the memory address range that maps to CXL and feed them to our counter implementation. On the other side, each instance of CXL type 3 device can connect to this server and request hotness monitoring services + provide parameters etc. Elements such as list threshold management and overflow detection etc are in the CXL HMU QEMU device mode. As noted above, we have an alternative approach that can closely couple things, so the device model registers the plugin directly and there is no server.

How to use it!
--------------

It runs a little slow but it runs and generates somewhat plausible outputs. I'd definitely suggest running it with the pass through optimization patch on the CXL staging tree (and a single direct connected device). Your millage will vary if you try to use other parameters, or hotness units beyond the first one (implementation far from complete!)

To run start the server in contrib/hmu/ providing a port number to listen on.

./chmu 4443

Then launch QEMU with something like the following.

qemu-system-aarch64 -icount shift=1 \
-plugin ../qemu/bin/native/contrib/plugins/libcache.so,port=4443,missfilterbase=1099511627776,missfiltersize=1099511627776,dcachesize=8192,dassoc=4,dblksize=64,icachesize=8192,iassoc=4,iblksize=64,l2cachesize=32768,l2assoc=16,l2blksize=64 \
-M virt,ras=on,nvdimm=on,gic-version=3,cxl=on,hmat=on -m 4g,maxmem=8g,slots=4 -cpu max -smp 4 \
-kernel Image \
-drive if=none,file=full.qcow2,format=qcow2,id=hd \
-device pcie-root-port,id=root_port1 \
-device virtio-blk-pci,drive=hd,x-max-bounce-buffer-size=512k \
-nographic -no-reboot -append 'earlycon memblock=debug root=/dev/vda2 fsck.mode=skip maxcpus=4 tp_printk' \
-monitor telnet:127.0.0.1:1234,server,nowait -bios QEMU_EFI.fd \
-object memory-backend-ram,size=4G,id=mem0 \
-object memory-backend-file,id=cxl-mem1,share=on,mem-path=/tmp/t3_cxl1.raw,size=1G,align=256M \
-object memory-backend-file,id=cxl-mem2,share=on,mem-path=/tmp/t3_cxl2.raw,size=1G,align=256M \
-object memory-backend-file,id=cxl-lsa1,share=on,mem-path=/tmp/t3_lsa1.raw,size=1M,align=1M \
-object memory-backend-file,id=cxl-mem3,share=on,mem-path=/tmp/t3_cxl3.raw,size=1G,align=256M \
-object memory-backend-file,id=cxl-mem4,share=on,mem-path=/tmp/t3_cxl4.raw,size=1G,align=256M \
-object memory-backend-file,id=cxl-lsa2,share=on,mem-path=/tmp/t3_lsa2.raw,size=1M,align=1M \
-device pxb-cxl,bus_nr=12,bus=pcie.0,id=cxl.1,hdm_for_passthrough=true,numa_node=0\
-device cxl-rp,port=0,bus=cxl.1,id=cxl_rp_port0,chassis=0,slot=2 \
-device cxl-type3,bus=cxl_rp_port0,volatile-memdev=cxl-mem1,id=cxl-pmem1,lsa=cxl-lsa1,sn=3,x-speed=32,x-width=16,chmu-port=4443 \
-machine cxl-fmw.0.targets.0=cxl.1,cxl-fmw.0.size=8G,cxl-fmw.0.interleave-granularity=1k \
-numa node,nodeid=0,cpus=0-3,memdev=mem0 \
-numa node,nodeid=1 \
-object acpi-generic-initiator,id=bob2,pci-dev=bob,node=1 \
-numa node,nodeid=2 \
-object acpi-generic-port,id=bob11,pci-bus=cxl.1,node=2 \

In the guest, create and bind the region - this brings up the CXL memory device so accesses go to the memory.

  cd /sys/bus/cxl/devices/decoder0.0/
  cat create_ram_region
  echo region0 > create_ram_region
  echo ram > /sys/bus/cxl/devices/decoder2.0/mode
  echo ram > /sys/bus/cxl/devices/decoder3.0/mode
  echo $((256 << 21)) > /sys/bus/cxl/devices/decoder2.0/dpa_size
  cd /sys/bus/cxl/devices/region0/
  echo 256 > interleave_granularity
  echo 1 > interleave_ways
  echo $((256 << 21)) > size
  echo decoder2.0 > target0
  echo 1 > commit
  echo region0 > /sys/bus/cxl/drivers/cxl_region/bind

Finally start perf with something like:

./perf record -a -e cxl_hmu_mem0.0.0/epoch_type=0,access_type=6,\
hotness_threshold=635,epoch_multiplier=4,epoch_scale=4,\
range_base=0,range_size=4096/ ./stress.sh

where stress.sh is

  sleep 2
  numactl --membind 3 stress-ng --vm 1 --vm-bytes 1M --vm-keep -t 5s
  sleep 2

See the results with
./perf report --dump-raw-trace | grep -A 200 HMU

Enjoy and have a good weekend.

Thanks,

Jonathan

Jonathan Cameron (3):
hw/cxl: Initial CXL Hotness Monitoring Unit Emulation
plugins: Add cache miss reporting over a socket.
contrib: Add example hotness monitoring unit server

 include/hw/cxl/cxl.h        |   1 +
 include/hw/cxl/cxl_chmu.h   | 154 ++++++++++++
 include/hw/cxl/cxl_device.h |  13 +-
 include/hw/cxl/cxl_pci.h    |   7 +-
 contrib/hmu/hmu.c           | 312 ++++++++++++++++++++++++
 contrib/plugins/cache.c     |  75 +++++-
 hw/cxl/cxl-chmu.c           | 459 ++++++++++++++++++++++++++++++++++++
 hw/mem/cxl_type3.c          |  25 +-
 hw/cxl/meson.build          |   1 +
 9 files changed, 1035 insertions(+), 12 deletions(-)
 create mode 100644 include/hw/cxl/cxl_chmu.h
 create mode 100644 contrib/hmu/hmu.c
 create mode 100644 hw/cxl/cxl-chmu.c
Signature · 1 line
2.43.0
#1 lore raw

RFC PATCH 1/3 hw/cxl: Initial CXL Hotness Monitoring Unit Emulation

Intended to support enabling in kernel. For now this is dumb and the data made up. That will change in the near future.

Instantiates 3 instances within one CHMU with separate interrupts.

---
 include/hw/cxl/cxl.h        |   1 +
 include/hw/cxl/cxl_chmu.h   | 154 ++++++++++++
 include/hw/cxl/cxl_device.h |  13 +-
 include/hw/cxl/cxl_pci.h    |   7 +-
 hw/cxl/cxl-chmu.c           | 459 ++++++++++++++++++++++++++++++++++++
 hw/mem/cxl_type3.c          |  25 +-
 hw/cxl/meson.build          |   1 +
 7 files changed, 655 insertions(+), 5 deletions(-)
include/hw/cxl/cxl.h +1 -0
index 857fa61898..bef856485f 100644
@@ -16,6 +16,7 @@
 #include "hw/pci/pci_host.h"
 #include "cxl_pci.h"
 #include "cxl_component.h"
+#include "cxl_chmu.h"
 #include "cxl_cpmu.h"
 #include "cxl_device.h"
 
include/hw/cxl/cxl_chmu.h new file +154 -0

154 lines not on this page · raw message

include/hw/cxl/cxl_device.h +12 -1
index 04c93cd753..f855cd69d9 100644
@@ -15,6 +15,7 @@
 #include "hw/register.h"
 #include "hw/cxl/cxl_events.h"
 
+#include "hw/cxl/cxl_chmu.h"
 #include "hw/cxl/cxl_cpmu.h"
 /*
  * The following is how a CXL device's Memory Device registers are laid out.
@@ -109,12 +110,20 @@
                   (x) * (1 << 16),                                      \
                   1 << 16)
 
+#define CXL_NUM_CHMU_INSTANCES 1
+#define CXL_CHMU_OFFSET(x)                                               \
+    QEMU_ALIGN_UP(CXL_MEMORY_DEVICE_REGISTERS_OFFSET +                  \
+                  CXL_MEMORY_DEVICE_REGISTERS_LENGTH +                  \
+                  (1 << 16) * CXL_NUM_CPMU_INSTANCES,                   \
+                  1 << 16)
+
 #define CXL_MMIO_SIZE                                                   \
     QEMU_ALIGN_UP(CXL_DEVICE_CAP_REG_SIZE +                             \
                   CXL_DEVICE_STATUS_REGISTERS_LENGTH +                  \
                   CXL_MAILBOX_REGISTERS_LENGTH +                        \
                   CXL_MEMORY_DEVICE_REGISTERS_LENGTH +                  \
-                  CXL_NUM_CPMU_INSTANCES * (1 << 16),                   \
+                  CXL_NUM_CPMU_INSTANCES * (1 << 16) +                  \
+                  CXL_NUM_CHMU_INSTANCES * (1 << 16),                   \
                   (1 << 16))
 
 /* CXL r3.1 Table 8-34: Command Return Codes */
@@ -231,6 +240,7 @@ typedef struct CXLCCI {
 typedef struct cxl_device_state {
     MemoryRegion device_registers;
     MemoryRegion cpmu_registers[CXL_NUM_CPMU_INSTANCES];
+    MemoryRegion chmu_registers[1];
     /* CXL r3.1 Section 8.2.8.3: Device Status Registers */
     struct {
         MemoryRegion device;
@@ -280,6 +290,7 @@ typedef struct cxl_device_state {
 
     const struct cxl_cmd (*cxl_cmd_set)[256];
     CPMUState cpmu[CXL_NUM_CPMU_INSTANCES];
+    CHMUState chmu[1];
     CXLEventLog event_logs[CXL_EVENT_TYPE_MAX];
 } CXLDeviceState;
 
include/hw/cxl/cxl_pci.h +4 -3
index c54ed54a25..88a5e3958e 100644
@@ -32,7 +32,7 @@
 #define PCIE_CXL3_FLEXBUS_PORT_DVSEC_LENGTH 0x20
 #define PCIE_CXL3_FLEXBUS_PORT_DVSEC_REVID  2
 
-#define REG_LOC_DVSEC_LENGTH 0x2c
+#define REG_LOC_DVSEC_LENGTH 0x34
 #define REG_LOC_DVSEC_REVID  0
 
 enum {
@@ -172,9 +172,9 @@ typedef struct CXLDVSECRegisterLocator {
     struct {
             uint32_t lo;
             uint32_t hi;
-    } reg_base[4];
+    } reg_base[5];
 } QEMU_PACKED CXLDVSECRegisterLocator;
-QEMU_BUILD_BUG_ON(sizeof(CXLDVSECRegisterLocator) != 0x2C);
+QEMU_BUILD_BUG_ON(sizeof(CXLDVSECRegisterLocator) != 0x34);
 
 /* BAR Equivalence Indicator */
 #define BEI_BAR_10H 0
@@ -190,5 +190,6 @@ QEMU_BUILD_BUG_ON(sizeof(CXLDVSECRegisterLocator) != 0x2C);
 #define RBI_BAR_VIRT_ACL   (2 << 8)
 #define RBI_CXL_DEVICE_REG (3 << 8)
 #define RBI_CXL_CPMU_REG   (4 << 8)
+#define RBI_CXL_CHMU_REG   (5 << 8)
 
 #endif
hw/cxl/cxl-chmu.c new file +459 -0

459 lines not on this page · raw message

hw/mem/cxl_type3.c +24 -1
index c1004ddae8..78426758af 100644
@@ -38,7 +38,10 @@ enum CXL_T3_MSIX_VECTOR {
     CXL_T3_MSIX_CPMU0,
     CXL_T3_MSIX_CPMU1,
     CXL_T3_MSIX_PCIE_DOE_COMPLIANCE,
-    CXL_T3_MSIX_VECTOR_NR
+    CXL_T3_MSIX_CHMU0_BASE,
+    /* One interrupt per CMUH instance in the block */
+    CXL_T3_MSIX_VECTOR_NR =
+        CXL_T3_MSIX_CHMU0_BASE + CXL_CHMU_INSTANCES_PER_BLOCK,
 };
 
 #define DWORD_BYTE 4
@@ -499,6 +502,8 @@ static void build_dvsecs(CXLType3Dev *ct3d)
             RBI_CXL_CPMU_REG | CXL_DEVICE_REG_BAR_IDX;
         regloc_dvsec->reg_base[2 + i].hi = 0;
     }
+    regloc_dvsec->reg_base[4].lo = CXL_CHMU_OFFSET(0) | RBI_CXL_CHMU_REG |
+        CXL_DEVICE_REG_BAR_IDX;
     cxl_component_create_dvsec(cxl_cstate, CXL2_TYPE3_DEVICE,
                                REG_LOC_DVSEC_LENGTH, REG_LOC_DVSEC,
                                REG_LOC_DVSEC_REVID, (uint8_t *)regloc_dvsec);
@@ -535,6 +540,17 @@ static void hdm_decoder_commit(CXLType3Dev *ct3d, int which)
     ctrl = FIELD_DP32(ctrl, CXL_HDM_DECODER0_CTRL, COMMITTED, 1);
 
     stl_le_p(cache_mem + R_CXL_HDM_DECODER0_CTRL + which * hdm_inc, ctrl);
+
+    if (which == 0) {
+        uint32_t low, high;
+        low = ldl_le_p(cache_mem + R_CXL_HDM_DECODER0_BASE_LO);
+        high = ldl_le_p(cache_mem + R_CXL_HDM_DECODER0_BASE_HI);
+        ct3d->cxl_dstate.chmu[0].base = ((uint64_t)high << 32) | (low & 0xf0000000);
+        
+        low = ldl_le_p(cache_mem + R_CXL_HDM_DECODER0_SIZE_LO);
+        high = ldl_le_p(cache_mem + R_CXL_HDM_DECODER0_SIZE_HI);
+        ct3d->cxl_dstate.chmu[0].size = ((uint64_t)high << 32) | (low & 0xf0000000);
+    }
 }
 
 static void hdm_decoder_uncommit(CXLType3Dev *ct3d, int which)
@@ -1008,6 +1024,12 @@ static void ct3_realize(PCIDevice *pci_dev, Error **errp)
                                  CXL_T3_MSIX_CPMU0);
     cxl_cpmu_register_block_init(OBJECT(pci_dev), &ct3d->cxl_dstate, 1,
                                  CXL_T3_MSIX_CPMU1);
+    rc = cxl_chmu_register_block_init(OBJECT(pci_dev), &ct3d->cxl_dstate,
+                                      0, CXL_T3_MSIX_CHMU0_BASE, errp);
+    if (rc) {
+        goto err_free_special_ops;
+    }
+
     pci_register_bar(pci_dev, CXL_DEVICE_REG_BAR_IDX,
                      PCI_BASE_ADDRESS_SPACE_MEMORY |
                          PCI_BASE_ADDRESS_MEM_TYPE_64,
@@ -1317,6 +1339,7 @@ static const Property ct3_props[] = {
                                 speed, PCIE_LINK_SPEED_32),
     DEFINE_PROP_PCIE_LINK_WIDTH("x-width", CXLType3Dev,
                                 width, PCIE_LINK_WIDTH_16),
+    DEFINE_PROP_UINT16("chmu-port", CXLType3Dev, cxl_dstate.chmu[0].port, 0), 
 };
 
 static uint64_t get_lsa_size(CXLType3Dev *ct3d)
hw/cxl/meson.build +1 -0
index 4db7cad267..c97e64b586 100644
@@ -6,6 +6,7 @@ system_ss.add(when: 'CONFIG_CXL',
                    'cxl-host.c',
                    'cxl-cdat.c',
                    'cxl-events.c',
+                   'cxl-chmu.c',
                    'cxl-cpmu.c',
                    'switch-mailbox-cci.c',
                ),
Signature · 1 line
2.43.0
#2 lore raw

RFC PATCH 2/3 plugins: Add cache miss reporting over a socket.

In reply to Jonathan Cameron

This allows an external program to act as a hotness tracker.

---
 contrib/plugins/cache.c | 75 +++++++++++++++++++++++++++++++++++++----
 1 file changed, 68 insertions(+), 7 deletions(-)
contrib/plugins/cache.c +68 -7
index 7baff86860..5af1e6559c 100644
@@ -7,10 +7,17 @@
 
 #include <inttypes.h>
 #include <stdio.h>
+#include <unistd.h>
 #include <glib.h>
+#include <sys/socket.h>
+#include <arpa/inet.h>
 
 #include <qemu-plugin.h>
 
+static int client_socket = -1;
+static uint64_t missfilterbase;
+static uint64_t missfiltersize;
+
 #define STRTOLL(x) g_ascii_strtoll(x, NULL, 10)
 
 QEMU_PLUGIN_EXPORT int qemu_plugin_version = QEMU_PLUGIN_VERSION;
@@ -104,6 +111,7 @@ static Cache **l2_ucaches;
 static GMutex *l1_dcache_locks;
 static GMutex *l1_icache_locks;
 static GMutex *l2_ucache_locks;
+static GMutex *socket_lock;
 
 static uint64_t l1_dmem_accesses;
 static uint64_t l1_imem_accesses;
@@ -385,6 +393,21 @@ static bool access_cache(Cache *cache, uint64_t addr)
     return false;
 }
 
+static void miss(uint64_t paddr)
+{
+    if (client_socket < 0) {
+        return;
+    }
+
+    if (paddr < missfilterbase || paddr >= missfilterbase + missfiltersize) {
+        return;
+    }
+
+    g_mutex_lock(socket_lock);
+    send(client_socket, &paddr, sizeof(paddr), 0);
+    g_mutex_unlock(socket_lock);
+}
+
 static void vcpu_mem_access(unsigned int vcpu_index, qemu_plugin_meminfo_t info,
                             uint64_t vaddr, void *userdata)
 {
@@ -395,9 +418,6 @@ static void vcpu_mem_access(unsigned int vcpu_index, qemu_plugin_meminfo_t info,
     bool hit_in_l1;
 
     hwaddr = qemu_plugin_get_hwaddr(info, vaddr);
-    if (hwaddr && qemu_plugin_hwaddr_is_io(hwaddr)) {
-        return;
-    }
 
     effective_addr = hwaddr ? qemu_plugin_hwaddr_phys_addr(hwaddr) : vaddr;
     cache_idx = vcpu_index % cores;
@@ -412,7 +432,11 @@ static void vcpu_mem_access(unsigned int vcpu_index, qemu_plugin_meminfo_t info,
     l1_dcaches[cache_idx]->accesses++;
     g_mutex_unlock(&l1_dcache_locks[cache_idx]);
 
-    if (hit_in_l1 || !use_l2) {
+    if (hit_in_l1) {
+        return;
+    }
+    if (!use_l2) {
+        miss(effective_addr);
         /* No need to access L2 */
         return;
     }
@@ -422,6 +446,7 @@ static void vcpu_mem_access(unsigned int vcpu_index, qemu_plugin_meminfo_t info,
         insn = userdata;
         __atomic_fetch_add(&insn->l2_misses, 1, __ATOMIC_SEQ_CST);
         l2_ucaches[cache_idx]->misses++;
+        miss(effective_addr);
     }
     l2_ucaches[cache_idx]->accesses++;
     g_mutex_unlock(&l2_ucache_locks[cache_idx]);
@@ -447,8 +472,12 @@ static void vcpu_insn_exec(unsigned int vcpu_index, void *userdata)
     l1_icaches[cache_idx]->accesses++;
     g_mutex_unlock(&l1_icache_locks[cache_idx]);
 
-    if (hit_in_l1 || !use_l2) {
-        /* No need to access L2 */
+    if (hit_in_l1) {
+        return;
+    }
+
+    if (!use_l2) {
+        miss(insn_addr);
         return;
     }
 
@@ -739,14 +768,16 @@ QEMU_PLUGIN_EXPORT
 int qemu_plugin_install(qemu_plugin_id_t id, const qemu_info_t *info,
                         int argc, char **argv)
 {
-    int i;
+    int i, port;
     int l1_iassoc, l1_iblksize, l1_icachesize;
     int l1_dassoc, l1_dblksize, l1_dcachesize;
     int l2_assoc, l2_blksize, l2_cachesize;
+    struct sockaddr_in server_addr;
 
     limit = 32;
     sys = info->system_emulation;
 
+    port = -1;
     l1_dassoc = 8;
     l1_dblksize = 64;
     l1_dcachesize = l1_dblksize * l1_dassoc * 32;
@@ -808,11 +839,39 @@ int qemu_plugin_install(qemu_plugin_id_t id, const qemu_info_t *info,
                 fprintf(stderr, "invalid eviction policy: %s\n", opt);
                 return -1;
             }
+        } else if (g_strcmp0(tokens[0], "port") == 0) {
+            port = STRTOLL(tokens[1]);
+        } else if (g_strcmp0(tokens[0], "missfilterbase") == 0) {
+            missfilterbase = STRTOLL(tokens[1]);
+        } else if (g_strcmp0(tokens[0], "missfiltersize") == 0) {
+            missfiltersize = STRTOLL(tokens[1]);
         } else {
             fprintf(stderr, "option parsing failed: %s\n", opt);
             return -1;
         }
     }
+    if (port >= -1) {
+        uint64_t paddr = 42; /* hello, I'm a provider */
+        client_socket = socket(AF_INET, SOCK_STREAM, IPPROTO_TCP);
+        if (client_socket < 0) {
+            printf("failed to create a socket\n");
+            return -1;
+        }
+        printf("Cache miss reported on on %lx size %lx\n",
+               missfilterbase, missfiltersize);
+        memset((char *)&server_addr, 0, sizeof(server_addr));
+        server_addr.sin_family = AF_INET;
+        server_addr.sin_addr.s_addr = htonl(INADDR_LOOPBACK);
+        server_addr.sin_port = htons(port);
+
+        if (connect(client_socket, (struct sockaddr *)&server_addr,
+                    sizeof(server_addr)) < 0) {
+            close(client_socket);
+            return -1;
+        }
+        /* Let it know we are a data provider */
+        send(client_socket, &paddr, sizeof(paddr), 0);
+    }
 
     policy_init();
 
@@ -840,6 +899,8 @@ int qemu_plugin_install(qemu_plugin_id_t id, const qemu_info_t *info,
         return -1;
     }
 
+    socket_lock = g_new0(GMutex, 1);
+
     l1_dcache_locks = g_new0(GMutex, cores);
     l1_icache_locks = g_new0(GMutex, cores);
     l2_ucache_locks = use_l2 ? g_new0(GMutex, cores) : NULL;
Signature · 1 line
2.43.0
#3 lore raw

Re: [RFC PATCH QEMU 2/3] plugins: Add cache miss reporting over a socket.

Quoted 150 lines · Jonathan Cameron

This allows an external program to act as a hotness tracker.

---
 contrib/plugins/cache.c | 75 +++++++++++++++++++++++++++++++++++++----
 1 file changed, 68 insertions(+), 7 deletions(-)
contrib/plugins/cache.c +56 -7
index 7baff86860..5af1e6559c 100644
@@ -7,10 +7,17 @@
 
 #include <inttypes.h>
 #include <stdio.h>
+#include <unistd.h>
 #include <glib.h>
+#include <sys/socket.h>
+#include <arpa/inet.h>
 
 #include <qemu-plugin.h>
 
+static int client_socket = -1;
+static uint64_t missfilterbase;
+static uint64_t missfiltersize;
+
 #define STRTOLL(x) g_ascii_strtoll(x, NULL, 10)
 
 QEMU_PLUGIN_EXPORT int qemu_plugin_version = QEMU_PLUGIN_VERSION;
@@ -104,6 +111,7 @@ static Cache **l2_ucaches;
 static GMutex *l1_dcache_locks;
 static GMutex *l1_icache_locks;
 static GMutex *l2_ucache_locks;
+static GMutex *socket_lock;
 
 static uint64_t l1_dmem_accesses;
 static uint64_t l1_imem_accesses;
@@ -385,6 +393,21 @@ static bool access_cache(Cache *cache, uint64_t addr)
     return false;
 }
 
+static void miss(uint64_t paddr)
+{
+    if (client_socket < 0) {
+        return;
+    }
+
+    if (paddr < missfilterbase || paddr >= missfilterbase + missfiltersize) {
+        return;
+    }
+
+    g_mutex_lock(socket_lock);
+    send(client_socket, &paddr, sizeof(paddr), 0);
+    g_mutex_unlock(socket_lock);
+}
+
 static void vcpu_mem_access(unsigned int vcpu_index, qemu_plugin_meminfo_t info,
                             uint64_t vaddr, void *userdata)
 {
@@ -395,9 +418,6 @@ static void vcpu_mem_access(unsigned int vcpu_index, qemu_plugin_meminfo_t info,
     bool hit_in_l1;
 
     hwaddr = qemu_plugin_get_hwaddr(info, vaddr);
-    if (hwaddr && qemu_plugin_hwaddr_is_io(hwaddr)) {
-        return;
-    }
 
     effective_addr = hwaddr ? qemu_plugin_hwaddr_phys_addr(hwaddr) : vaddr;
     cache_idx = vcpu_index % cores;
@@ -412,7 +432,11 @@ static void vcpu_mem_access(unsigned int vcpu_index, qemu_plugin_meminfo_t info,
     l1_dcaches[cache_idx]->accesses++;
     g_mutex_unlock(&l1_dcache_locks[cache_idx]);
 
-    if (hit_in_l1 || !use_l2) {
+    if (hit_in_l1) {
+        return;
+    }
+    if (!use_l2) {
+        miss(effective_addr);
         /* No need to access L2 */
         return;
     }
@@ -422,6 +446,7 @@ static void vcpu_mem_access(unsigned int vcpu_index, qemu_plugin_meminfo_t info,
         insn = userdata;
         __atomic_fetch_add(&insn->l2_misses, 1, __ATOMIC_SEQ_CST);
         l2_ucaches[cache_idx]->misses++;
+        miss(effective_addr);
     }
     l2_ucaches[cache_idx]->accesses++;
     g_mutex_unlock(&l2_ucache_locks[cache_idx]);
@@ -447,8 +472,12 @@ static void vcpu_insn_exec(unsigned int vcpu_index, void *userdata)
     l1_icaches[cache_idx]->accesses++;
     g_mutex_unlock(&l1_icache_locks[cache_idx]);
 
-    if (hit_in_l1 || !use_l2) {
-        /* No need to access L2 */
+    if (hit_in_l1) {
+        return;
+    }
+
+    if (!use_l2) {
+        miss(insn_addr);
         return;
     }
 
@@ -739,14 +768,16 @@ QEMU_PLUGIN_EXPORT
 int qemu_plugin_install(qemu_plugin_id_t id, const qemu_info_t *info,
                         int argc, char **argv)
 {
-    int i;
+    int i, port;
     int l1_iassoc, l1_iblksize, l1_icachesize;
     int l1_dassoc, l1_dblksize, l1_dcachesize;
     int l2_assoc, l2_blksize, l2_cachesize;
+    struct sockaddr_in server_addr;
 
     limit = 32;
     sys = info->system_emulation;
 
+    port = -1;
     l1_dassoc = 8;
     l1_dblksize = 64;
     l1_dcachesize = l1_dblksize * l1_dassoc * 32;
@@ -808,11 +839,39 @@ int qemu_plugin_install(qemu_plugin_id_t id, const qemu_info_t *info,
                 fprintf(stderr, "invalid eviction policy: %s\n", opt);
                 return -1;
             }
+        } else if (g_strcmp0(tokens[0], "port") == 0) {
+            port = STRTOLL(tokens[1]);
+        } else if (g_strcmp0(tokens[0], "missfilterbase") == 0) {
+            missfilterbase = STRTOLL(tokens[1]);
+        } else if (g_strcmp0(tokens[0], "missfiltersize") == 0) {
+            missfiltersize = STRTOLL(tokens[1]);
         } else {
             fprintf(stderr, "option parsing failed: %s\n", opt);
             return -1;
         }
     }
+    if (port >= -1) {
+        uint64_t paddr = 42; /* hello, I'm a provider */
+        client_socket = socket(AF_INET, SOCK_STREAM, IPPROTO_TCP);
+        if (client_socket < 0) {
+            printf("failed to create a socket\n");
+            return -1;
+        }
+        printf("Cache miss reported on on %lx size %lx\n",
+               missfilterbase, missfiltersize);
+        memset((char *)&server_addr, 0, sizeof(server_addr));
+        server_addr.sin_family = AF_INET;
+        server_addr.sin_addr.s_addr = htonl(INADDR_LOOPBACK);

contrib/plugins/cache.c int qemu_plugin_install(qemu_plugin_id_t id, const qemu_info_t *info, · line 874

Quoted 10 lines
contrib/plugins/cache.c +10 -0
+        server_addr.sin_port = htons(port);
+
+        if (connect(client_socket, (struct sockaddr *)&server_addr,
+                    sizeof(server_addr)) < 0) {
+            close(client_socket);
+            return -1;
+        }
+        /* Let it know we are a data provider */
+        send(client_socket, &paddr, sizeof(paddr), 0);
+    }

No particular objections to the patch as is. I do wonder if it would be worth exposing a chardev pipe to plugins so we could take advantage of QEMU's flexible redirection handling.

But not a blocker for this.

Quoted 12 lines
contrib/plugins/cache.c +2 -0
     policy_init();
 
@@ -840,6 +899,8 @@ int qemu_plugin_install(qemu_plugin_id_t id, const qemu_info_t *info,
         return -1;
     }
 
+    socket_lock = g_new0(GMutex, 1);
+
     l1_dcache_locks = g_new0(GMutex, cores);
     l1_icache_locks = g_new0(GMutex, cores);
     l2_ucache_locks = use_l2 ? g_new0(GMutex, cores) : NULL;
Signature · 2 lines
Alex Bennée
Virtualisation Tech Lead @ Linaro
#4 lore raw

RFC PATCH 1/1 contrib: Add example hotness monitoring unit server

In reply to Jonathan Cameron

This is used inconjuction with the cache plugin (with port parameter supplied) and the CXL Type 3 device with a hotness monitoring unit (chmu-port parameter supplied).

It implements a very basic oracle with a counter per 4KiB page and simple loop to find large counts. The hotlist length is controlled by the QEMU device implementation.

This is only responsible for the data handling events etc are a problem for the CXL HMU emulation.

Note that when running this things are fairly slow.

---
 contrib/hmu/hmu.c | 312 ++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 312 insertions(+)
contrib/hmu/hmu.c new file +312 -0
new file mode 100644
index 0000000000..aa47efd98b
@@ -0,0 +1,312 @@
+#include <stdio.h>
+#include <string.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include <stdbool.h>
+#include <stdint.h>
+#include <pthread.h>
+#include <arpa/inet.h>
+
+#define ID_PROVIDER 42
+#define ID_CONSUMER 41
+
+/* Move to shared header */
+enum consumer_request {
+    QUERY_TAIL,
+    QUERY_HEAD,
+    SET_HEAD,
+    SET_HOTLIST_SIZE,
+    QUERY_HOTLIST_ENTRY,
+    SIGNAL_EPOCH_END,
+    SET_ENABLED,
+    SET_NUMBER_GRANUALS,
+    SET_HPA_BASE,
+    SET_HPA_SIZE,
+};
+
+struct tracking_instance {
+    uint64_t base, size;
+    uint16_t head, tail;
+    uint16_t hotlist_length;
+    uint64_t *hotlist;
+    int32_t *counters;
+    size_t num_counters;
+    bool enabled;
+};
+
+#define MAX_INSTANCES 16
+static int num_tracking_instances;
+static struct tracking_instance *instances[MAX_INSTANCES] = {};
+/*
+ * Instances never removed so this only protects the index against
+ * parallel creations.
+ */
+pthread_mutex_t instances_lock;
+static int register_tracker(struct tracking_instance *inst)
+{
+    pthread_mutex_lock(&instances_lock);
+    if (num_tracking_instances >= MAX_INSTANCES) {
+        pthread_mutex_unlock(&instances_lock);
+        return -1;
+    }
+    instances[num_tracking_instances++] = inst;
+    printf("registered %d\n", num_tracking_instances);
+    pthread_mutex_unlock(&instances_lock);
+    return 0;
+}
+
+static void notify_tracker(struct tracking_instance *inst, uint64_t paddr)
+{
+    uint64_t offset;
+
+    if (paddr < inst->base || paddr >= inst->base + inst->size) {
+        return;
+    }
+    /* Fixme: multiple regions */
+    offset = (paddr - inst->base) / 4096;
+
+    /*  TODO - check masking */
+
+    if (!inst->counters) {
+        printf("No counter storage\n");
+        return;
+    }
+    if (offset >= inst->num_counters) {
+        printf("out of range? %lx %lx\n", offset, inst->num_counters);
+        return;
+    }
+    inst->counters[offset]++;
+}
+
+/* CHMU instance in QEMU */
+static void *provider_innerloop(void * _socket)
+{
+    int socket = *(int *)_socket;
+    uint64_t paddr;
+    int rc;
+
+    printf("Provider connected\n");
+    while (1) {
+        rc = read(socket, &paddr, sizeof(paddr));
+        if (rc == 0) {
+            return NULL;
+        }
+       /* Lock not taken as instances only goes up which should be safe */
+        for (int i = 0; i < num_tracking_instances; i++)
+            if (instances[i]->enabled) {
+                notify_tracker(instances[i], paddr);
+            }
+    }
+}
+
+
+/* Cache plugin hopefully squirting us some data */
+static void *consumer_innerloop(void *_socket)
+{
+    int socket = *(int *)_socket;
+    /* for now all chmu have 3 instances */
+    struct tracking_instance insts[3] = {};
+    /* Instance, command, parameter */
+    uint64_t paddr[3];
+    int rc;
+
+    for (int i = 0; i < 3; i++) {
+        rc = register_tracker(&insts[i]);
+        if (rc) {
+            printf("Failed to register tracker\n");
+            return NULL;
+            /* todo cleanup to not have partial trackers registered */
+        }
+    }
+    printf("Consumer connected\n");
+
+    while (1) {
+        uint64_t reply, param;
+        enum consumer_request request;
+
+        struct tracking_instance *inst;
+
+        rc = read(socket, paddr, sizeof(paddr));
+        if (rc < sizeof(paddr)) {
+            printf("short message %x\n", rc);
+            return NULL;
+        }
+        if (paddr[0] > 3) {
+            printf("garbage\n");
+            exit(-1);
+        }
+        inst = &insts[paddr[0]];
+        request = paddr[1];
+        param = paddr[2];
+
+        switch (request) {
+        case QUERY_TAIL:
+            reply = inst->tail;
+            break;
+        case QUERY_HEAD:
+            reply = inst->head;
+            break;
+        case SET_HEAD:
+            reply = param;
+            inst->head = param;
+            break;
+        case SET_HOTLIST_SIZE: {
+            uint64_t *newlist;
+            reply = param;
+            inst->hotlist_length = param;
+            newlist = realloc(inst->hotlist, sizeof(*inst->hotlist) * param);
+            if (!newlist) {
+                printf("failed to allocate hotlist\n");
+                break;
+            }
+            inst->hotlist = newlist;
+            break;
+        }
+        case QUERY_HOTLIST_ENTRY:
+            if (param >= inst->hotlist_length) {
+                printf("out of range hotlist read?\n");
+                break;
+            }
+            reply = inst->hotlist[param];
+            break;
+        case SIGNAL_EPOCH_END: {
+            int space;
+            int added = 0;
+            printf("into epoch end\n");
+            reply = param;
+
+            if (insts->tail > inst->head) {
+                space = inst->tail - inst->head;
+            } else {
+                space = inst->hotlist_length - inst->tail +
+                    inst->head;
+            }
+            if (!inst->counters) {
+                printf("How did we reach end of an epoque without counters?\n");
+                break;
+            }
+            for (int i = 0; i < inst->num_counters; i++) {
+                if (!(inst->counters[i] > 0)) {
+                    continue;
+                }
+                inst->hotlist[inst->tail] =
+                    (uint64_t)inst->counters[i] | ((uint64_t)i << 32);
+                printf("added hotlist element %lx at %u\n",
+                       inst->hotlist[inst->tail], inst->tail);
+                inst->tail = (inst->tail + 1) % inst->hotlist_length;
+                added++;
+                if (added == space) {
+                    break;
+                }
+            }
+            memset(inst->counters, 0,
+                   inst->num_counters * sizeof(*inst->counters));
+
+            printf("End of epoch %u %u\n", inst->head, inst->tail);
+            /* Overflow hadnling based on fullness detection in qemu */
+            break;
+        }
+        case SET_ENABLED:
+            reply = param;
+            inst->enabled = !!param;
+            printf("enabled? %d\n", inst->enabled);
+            break;
+        case SET_NUMBER_GRANUALS: { /* FIXME Should derive from granual size */
+            uint32_t *newcounters;
+
+            reply = param;
+            newcounters = realloc(inst->counters,
+                                  sizeof(*inst->counters) *
+                                  param);
+            if (!newcounters) {
+                printf("Failed to allocate counter storage\n");
+            }
+            printf("allocated space for %lu counters\n", param);
+            inst->counters = newcounters;
+            inst->num_counters = param;
+            break;
+        }
+        case SET_HPA_BASE:
+            reply = param;
+            inst->base = param;
+            break;
+        case SET_HPA_SIZE: /* Size */
+            reply = param;
+            inst->size = param;
+            break;
+        default:
+            printf("No idea yet\n");
+            break;
+        }
+        write(socket, &reply, sizeof(reply));
+    }
+}
+
+int main(int argc, char **argv)
+{
+    int server_fd, new_socket;
+    struct sockaddr_in address;
+    int opt = 1;
+    int addrlen = sizeof(address);
+    uint64_t paddr;
+    unsigned short port;
+
+    if (argc < 2) {
+        printf("Please provide port to listen on\n");
+        return -1;
+    }
+    port = atoi(argv[1]);
+
+    if ((server_fd = socket(AF_INET, SOCK_STREAM, IPPROTO_TCP)) == 0) {
+        return -1;
+    }
+
+    if (setsockopt(server_fd, SOL_SOCKET, SO_REUSEADDR | SO_REUSEPORT,
+                   &opt, sizeof(opt))) {
+        return -1;
+    }
+    address.sin_family = AF_INET;
+    address.sin_addr.s_addr = INADDR_ANY;
+    address.sin_port = htons(port);
+
+    if (bind(server_fd, (struct sockaddr *)&address, sizeof(address)) < 0) {
+        return -1;
+    }
+
+    printf("Listening on port %u\n", port);
+    if (listen(server_fd, 3) < 0) {
+        return -1;
+    }
+
+    while (1) {
+        int rc;
+        pthread_t thread;
+        if ((new_socket = accept(server_fd, (struct sockaddr *)&address,
+                                 (socklen_t *)&addrlen)) < 0) {
+            exit(-1);
+        }
+
+        rc = read(new_socket, &paddr, sizeof(paddr));
+        if (rc == 0) {
+            return 0;
+        }
+
+        if (paddr == ID_PROVIDER) {
+            if (pthread_create(&thread, NULL, provider_innerloop,
+                               &new_socket)) {
+                printf("thread create fail\n");
+            };
+        } else if (paddr == ID_CONSUMER) {
+            if (pthread_create(&thread, NULL, consumer_innerloop,
+                               &new_socket)) {
+                printf("thread create fail\n");
+            };
+        } else {
+            printf("No idea what this was - initial value not provider or consumer\n");
+            close(new_socket);
+            return 0;
+        }
+    }
+
+    return 0;
+}
Signature · 1 line
2.43.0
#5 lore raw
In reply to Jonathan Cameron

Hi Jonathan,

thanks for posting this. It's a creative usage of plugins.

I think that your current approach, decoupling plugins, CHMU and device model is a good thing.

I'm not familiar with CXL, but one question that comes to my mind is:
Is that mandatory to do this analysis during execution (vs dumping binary traces from CHMU and plugin and running an analysis post execution)?

Regards,
Pierrick

Quoted 189 lines · Jonathan Cameron

Hi All,

This is an RFC mainly to seek feedback on the approach used, particularly the aspect of how to get data from a TCG plugin into a device model. Two options that we have tried
1. Socket over which the plugin sends data to an external server (as seen here)
2. Register and manage a plugin from within a device model

The external server approach keeps things loosely coupled, but at the cost of separately maintaining that server, protocol definitions etc and some overhead.
The closely couple solution is neater, but I suspect might be controversial (hence I didn't start with that :)

The code here is at best a PoC to illustrate what we have in mind It's not nice code at all, feature gaps, bugs and all! So whilst review is always welcome I'm not requesting it for now.

Kernel support was posted a while back but was done against fake data (still supported here if you don't provide the port parameter to the type3 device)
https://lore.kernel.org/linux-cxl/[email protected]/
I'll post a minor update of that driver shortly to take into account a few specification clarifications but it should work with this without those.

Note there are some other patches on the tree I generated this from so this may not apply to upstream. Easiest is probably to test using gitlab.com/jic23/qemu cxl-2025-01-24

Thanks to Niyas for his suggestions on how to make all this work!

Background
----------

What is the Compute eXpress Link Hotness Monitoring unit and what is it for?
- In a tiered memory equipped server with the slow tier being attached via CXL the expectation is a given workload will benefit from putting data that is frequently fetched from memory in lower latency directly attached DRAM. Less frequently used data can be served from the CXL attached memory with no significant loss of performance. Any data that is hot enough to almost always be in cache doesn't matter as it is only fetch from memory occasionally.
- Working out which memory is best places where is hard to do and in some workloads a dynamic problem. As such we need something we can measure to provide some indication of what data is in the wrong place. There are existing techniques to do this (page faulting, various CPU tracing systems, access bit scanning etc) but they all have significant overheads.
- Monitoring accesses on the CXL device provides a path to getting good data without those overheads. These units are known as CXL Hotness Monitoring Units or CHMUs. Loosely speaking they count accesses to granuals of data (e.g. 4KiB pages). Exactly how they do that and where they sacrifice data accuracy is an implementation trade off.

Why do we need a model that gives real data?
- In general there is a need to develop software on top of these units to move data to the right place. Hard to evaluate that if we are making up the info on what is 'hot'.
- Need to allow for a bunch of 'impdef' solutions. Note that CHMU in this patch set is an oracle - it has enough counters to count every access. That's not realistic but it doesn't get me shouted at by our architecture teams for giving away any secrets. If we move forward with this, I'll probably implement a limited counter + full CAM solution (also unrealistic, but closer to real) I'd be very interested in contributions of other approaches (there are lots in the literature, under the term top-k)
- Resources will be constrained, so whilst a CHMU might in theory allow monitoring everything at once, that will come with a big accuracy cost. We need to design the algorithms that give us good data given those constraints.

So we need a solution to explore the design space and develop the software to take advantage of this hardware (there are various LSF/MM proposals on how to use this an other ways of tracking hotness).
https://lore.kernel.org/all/[email protected]/
https://lore.kernel.org/lkml/Z4XUoWlU-UgRik18@gourry-fedora-PF4VCD3F/T/

QEMU plugins give us a way to do this. In particular the existing Cache plugin can be easily modified to tell use what memory addresses missed at the last level of emulated cache. We can then filter those for the memory address range that maps to CXL and feed them to our counter implementation. On the other side, each instance of CXL type 3 device can connect to this server and request hotness monitoring services + provide parameters etc. Elements such as list threshold management and overflow detection etc are in the CXL HMU QEMU device mode. As noted above, we have an alternative approach that can closely couple things, so the device model registers the plugin directly and there is no server.

How to use it!
--------------

It runs a little slow but it runs and generates somewhat plausible outputs. I'd definitely suggest running it with the pass through optimization patch on the CXL staging tree (and a single direct connected device). Your millage will vary if you try to use other parameters, or hotness units beyond the first one (implementation far from complete!)

To run start the server in contrib/hmu/ providing a port number to listen on.

./chmu 4443

Then launch QEMU with something like the following.

qemu-system-aarch64 -icount shift=1 \
-plugin ../qemu/bin/native/contrib/plugins/libcache.so,port=4443,missfilterbase=1099511627776,missfiltersize=1099511627776,dcachesize=8192,dassoc=4,dblksize=64,icachesize=8192,iassoc=4,iblksize=64,l2cachesize=32768,l2assoc=16,l2blksize=64 \
-M virt,ras=on,nvdimm=on,gic-version=3,cxl=on,hmat=on -m 4g,maxmem=8g,slots=4 -cpu max -smp 4 \
-kernel Image \
-drive if=none,file=full.qcow2,format=qcow2,id=hd \
-device pcie-root-port,id=root_port1 \
-device virtio-blk-pci,drive=hd,x-max-bounce-buffer-size=512k \
-nographic -no-reboot -append 'earlycon memblock=debug root=/dev/vda2 fsck.mode=skip maxcpus=4 tp_printk' \
-monitor telnet:127.0.0.1:1234,server,nowait -bios QEMU_EFI.fd \
-object memory-backend-ram,size=4G,id=mem0 \
-object memory-backend-file,id=cxl-mem1,share=on,mem-path=/tmp/t3_cxl1.raw,size=1G,align=256M \
-object memory-backend-file,id=cxl-mem2,share=on,mem-path=/tmp/t3_cxl2.raw,size=1G,align=256M \

  -object memory-backend-file,id=cxl-lsa1,share=on,mem-path=/tmp/t3_lsa1.raw,size=1M,align=1M \
   -object memory-backend-file,id=cxl-mem3,share=on,mem-path=/tmp/t3_cxl3.raw,size=1G,align=256M \

-object memory-backend-file,id=cxl-mem4,share=on,mem-path=/tmp/t3_cxl4.raw,size=1G,align=256M \
-object memory-backend-file,id=cxl-lsa2,share=on,mem-path=/tmp/t3_lsa2.raw,size=1M,align=1M \
-device pxb-cxl,bus_nr=12,bus=pcie.0,id=cxl.1,hdm_for_passthrough=true,numa_node=0\
-device cxl-rp,port=0,bus=cxl.1,id=cxl_rp_port0,chassis=0,slot=2 \
-device cxl-type3,bus=cxl_rp_port0,volatile-memdev=cxl-mem1,id=cxl-pmem1,lsa=cxl-lsa1,sn=3,x-speed=32,x-width=16,chmu-port=4443 \
-machine cxl-fmw.0.targets.0=cxl.1,cxl-fmw.0.size=8G,cxl-fmw.0.interleave-granularity=1k \
-numa node,nodeid=0,cpus=0-3,memdev=mem0 \
-numa node,nodeid=1 \
-object acpi-generic-initiator,id=bob2,pci-dev=bob,node=1 \
-numa node,nodeid=2 \
-object acpi-generic-port,id=bob11,pci-bus=cxl.1,node=2 \

In the guest, create and bind the region - this brings up the CXL memory device so accesses go to the memory.

   cd /sys/bus/cxl/devices/decoder0.0/
   cat create_ram_region
   echo region0 > create_ram_region
   echo ram > /sys/bus/cxl/devices/decoder2.0/mode
   echo ram > /sys/bus/cxl/devices/decoder3.0/mode
   echo $((256 << 21)) > /sys/bus/cxl/devices/decoder2.0/dpa_size
   cd /sys/bus/cxl/devices/region0/
   echo 256 > interleave_granularity
   echo 1 > interleave_ways
   echo $((256 << 21)) > size
   echo decoder2.0 > target0
   echo 1 > commit
   echo region0 > /sys/bus/cxl/drivers/cxl_region/bind

Finally start perf with something like:

./perf record -a -e cxl_hmu_mem0.0.0/epoch_type=0,access_type=6,\
hotness_threshold=635,epoch_multiplier=4,epoch_scale=4,\
range_base=0,range_size=4096/ ./stress.sh

where stress.sh is

   sleep 2
   numactl --membind 3 stress-ng --vm 1 --vm-bytes 1M --vm-keep -t 5s
   sleep 2

See the results with
./perf report --dump-raw-trace | grep -A 200 HMU

Enjoy and have a good weekend.

Thanks,

Jonathan

Jonathan Cameron (3):
hw/cxl: Initial CXL Hotness Monitoring Unit Emulation
plugins: Add cache miss reporting over a socket.
contrib: Add example hotness monitoring unit server

  include/hw/cxl/cxl.h        |   1 +
  include/hw/cxl/cxl_chmu.h   | 154 ++++++++++++
  include/hw/cxl/cxl_device.h |  13 +-
  include/hw/cxl/cxl_pci.h    |   7 +-
  contrib/hmu/hmu.c           | 312 ++++++++++++++++++++++++
  contrib/plugins/cache.c     |  75 +++++-
  hw/cxl/cxl-chmu.c           | 459 ++++++++++++++++++++++++++++++++++++
  hw/mem/cxl_type3.c          |  25 +-
  hw/cxl/meson.build          |   1 +
  9 files changed, 1035 insertions(+), 12 deletions(-)

create mode 100644 include/hw/cxl/cxl_chmu.h
create mode 100644 contrib/hmu/hmu.c
create mode 100644 hw/cxl/cxl-chmu.c

#6 lore raw
Quoted 10 lines · On Fri, 24 Jan 2025 12:55:52 -0800 Pierrick Bouvier…

Hi Jonathan,

thanks for posting this. It's a creative usage of plugins.

I think that your current approach, decoupling plugins, CHMU and device model is a good thing.

I'm not familiar with CXL, but one question that comes to my mind is:
Is that mandatory to do this analysis during execution (vs dumping binary traces from CHMU and plugin and running an analysis post execution)?

Short answer is that post run analysis isn't of much use for developing the OS software story. It works to some degree if you are designing the tracking hardware or algorithms to use that hardware capture a snapshot of hotness - dealing with lack of counters, that sort of thing.

The main intent of this support is to drive live usage of the data in the OS. So it gets this hotness information and migrates more frequently accessed memory to a 'nearer'/lower latency memory node.

From an OS point of view there will be two ways it uses it:
1) Offline application optimization - that aligns with your suggestion of offline analysis but would typically still need to be live because we have to do the reverse maps and work out what was allocated in particular locations. Not impossible to dump that information from QEMU + the guest OS but the usage flow would then look quite different from what makes sense on real hardware where all the data is available to the host OS directly.
2) Migration of memory. This will dynamically change the PA backing a VA whilst applications are running. The aim being to develop how that happens, we need the dynamic state.

Jonathan

Quoted 195 lines

Regards,
Pierrick

Quoted 189 lines · Jonathan Cameron

Hi All,

This is an RFC mainly to seek feedback on the approach used, particularly the aspect of how to get data from a TCG plugin into a device model. Two options that we have tried
1. Socket over which the plugin sends data to an external server (as seen here)
2. Register and manage a plugin from within a device model

The external server approach keeps things loosely coupled, but at the cost of separately maintaining that server, protocol definitions etc and some overhead.
The closely couple solution is neater, but I suspect might be controversial (hence I didn't start with that :)

The code here is at best a PoC to illustrate what we have in mind It's not nice code at all, feature gaps, bugs and all! So whilst review is always welcome I'm not requesting it for now.

Kernel support was posted a while back but was done against fake data (still supported here if you don't provide the port parameter to the type3 device)
https://lore.kernel.org/linux-cxl/[email protected]/
I'll post a minor update of that driver shortly to take into account a few specification clarifications but it should work with this without those.

Note there are some other patches on the tree I generated this from so this may not apply to upstream. Easiest is probably to test using gitlab.com/jic23/qemu cxl-2025-01-24

Thanks to Niyas for his suggestions on how to make all this work!

Background
----------

What is the Compute eXpress Link Hotness Monitoring unit and what is it for?
- In a tiered memory equipped server with the slow tier being attached via CXL the expectation is a given workload will benefit from putting data that is frequently fetched from memory in lower latency directly attached DRAM. Less frequently used data can be served from the CXL attached memory with no significant loss of performance. Any data that is hot enough to almost always be in cache doesn't matter as it is only fetch from memory occasionally.
- Working out which memory is best places where is hard to do and in some workloads a dynamic problem. As such we need something we can measure to provide some indication of what data is in the wrong place. There are existing techniques to do this (page faulting, various CPU tracing systems, access bit scanning etc) but they all have significant overheads.
- Monitoring accesses on the CXL device provides a path to getting good data without those overheads. These units are known as CXL Hotness Monitoring Units or CHMUs. Loosely speaking they count accesses to granuals of data (e.g. 4KiB pages). Exactly how they do that and where they sacrifice data accuracy is an implementation trade off.

Why do we need a model that gives real data?
- In general there is a need to develop software on top of these units to move data to the right place. Hard to evaluate that if we are making up the info on what is 'hot'.
- Need to allow for a bunch of 'impdef' solutions. Note that CHMU in this patch set is an oracle - it has enough counters to count every access. That's not realistic but it doesn't get me shouted at by our architecture teams for giving away any secrets. If we move forward with this, I'll probably implement a limited counter + full CAM solution (also unrealistic, but closer to real) I'd be very interested in contributions of other approaches (there are lots in the literature, under the term top-k)
- Resources will be constrained, so whilst a CHMU might in theory allow monitoring everything at once, that will come with a big accuracy cost. We need to design the algorithms that give us good data given those constraints.

So we need a solution to explore the design space and develop the software to take advantage of this hardware (there are various LSF/MM proposals on how to use this an other ways of tracking hotness).
https://lore.kernel.org/all/[email protected]/
https://lore.kernel.org/lkml/Z4XUoWlU-UgRik18@gourry-fedora-PF4VCD3F/T/

QEMU plugins give us a way to do this. In particular the existing Cache plugin can be easily modified to tell use what memory addresses missed at the last level of emulated cache. We can then filter those for the memory address range that maps to CXL and feed them to our counter implementation. On the other side, each instance of CXL type 3 device can connect to this server and request hotness monitoring services + provide parameters etc. Elements such as list threshold management and overflow detection etc are in the CXL HMU QEMU device mode. As noted above, we have an alternative approach that can closely couple things, so the device model registers the plugin directly and there is no server.

How to use it!
--------------

It runs a little slow but it runs and generates somewhat plausible outputs. I'd definitely suggest running it with the pass through optimization patch on the CXL staging tree (and a single direct connected device). Your millage will vary if you try to use other parameters, or hotness units beyond the first one (implementation far from complete!)

To run start the server in contrib/hmu/ providing a port number to listen on.

./chmu 4443

Then launch QEMU with something like the following.

qemu-system-aarch64 -icount shift=1 \
-plugin ../qemu/bin/native/contrib/plugins/libcache.so,port=4443,missfilterbase=1099511627776,missfiltersize=1099511627776,dcachesize=8192,dassoc=4,dblksize=64,icachesize=8192,iassoc=4,iblksize=64,l2cachesize=32768,l2assoc=16,l2blksize=64 \
-M virt,ras=on,nvdimm=on,gic-version=3,cxl=on,hmat=on -m 4g,maxmem=8g,slots=4 -cpu max -smp 4 \
-kernel Image \
-drive if=none,file=full.qcow2,format=qcow2,id=hd \
-device pcie-root-port,id=root_port1 \
-device virtio-blk-pci,drive=hd,x-max-bounce-buffer-size=512k \
-nographic -no-reboot -append 'earlycon memblock=debug root=/dev/vda2 fsck.mode=skip maxcpus=4 tp_printk' \
-monitor telnet:127.0.0.1:1234,server,nowait -bios QEMU_EFI.fd \
-object memory-backend-ram,size=4G,id=mem0 \
-object memory-backend-file,id=cxl-mem1,share=on,mem-path=/tmp/t3_cxl1.raw,size=1G,align=256M \
-object memory-backend-file,id=cxl-mem2,share=on,mem-path=/tmp/t3_cxl2.raw,size=1G,align=256M \

  -object memory-backend-file,id=cxl-lsa1,share=on,mem-path=/tmp/t3_lsa1.raw,size=1M,align=1M \
   -object memory-backend-file,id=cxl-mem3,share=on,mem-path=/tmp/t3_cxl3.raw,size=1G,align=256M \

-object memory-backend-file,id=cxl-mem4,share=on,mem-path=/tmp/t3_cxl4.raw,size=1G,align=256M \
-object memory-backend-file,id=cxl-lsa2,share=on,mem-path=/tmp/t3_lsa2.raw,size=1M,align=1M \
-device pxb-cxl,bus_nr=12,bus=pcie.0,id=cxl.1,hdm_for_passthrough=true,numa_node=0\
-device cxl-rp,port=0,bus=cxl.1,id=cxl_rp_port0,chassis=0,slot=2 \
-device cxl-type3,bus=cxl_rp_port0,volatile-memdev=cxl-mem1,id=cxl-pmem1,lsa=cxl-lsa1,sn=3,x-speed=32,x-width=16,chmu-port=4443 \
-machine cxl-fmw.0.targets.0=cxl.1,cxl-fmw.0.size=8G,cxl-fmw.0.interleave-granularity=1k \
-numa node,nodeid=0,cpus=0-3,memdev=mem0 \
-numa node,nodeid=1 \
-object acpi-generic-initiator,id=bob2,pci-dev=bob,node=1 \
-numa node,nodeid=2 \
-object acpi-generic-port,id=bob11,pci-bus=cxl.1,node=2 \

In the guest, create and bind the region - this brings up the CXL memory device so accesses go to the memory.

   cd /sys/bus/cxl/devices/decoder0.0/
   cat create_ram_region
   echo region0 > create_ram_region
   echo ram > /sys/bus/cxl/devices/decoder2.0/mode
   echo ram > /sys/bus/cxl/devices/decoder3.0/mode
   echo $((256 << 21)) > /sys/bus/cxl/devices/decoder2.0/dpa_size
   cd /sys/bus/cxl/devices/region0/
   echo 256 > interleave_granularity
   echo 1 > interleave_ways
   echo $((256 << 21)) > size
   echo decoder2.0 > target0
   echo 1 > commit
   echo region0 > /sys/bus/cxl/drivers/cxl_region/bind

Finally start perf with something like:

./perf record -a -e cxl_hmu_mem0.0.0/epoch_type=0,access_type=6,\
hotness_threshold=635,epoch_multiplier=4,epoch_scale=4,\
range_base=0,range_size=4096/ ./stress.sh

where stress.sh is

   sleep 2
   numactl --membind 3 stress-ng --vm 1 --vm-bytes 1M --vm-keep -t 5s
   sleep 2

See the results with
./perf report --dump-raw-trace | grep -A 200 HMU

Enjoy and have a good weekend.

Thanks,

Jonathan

Jonathan Cameron (3):
hw/cxl: Initial CXL Hotness Monitoring Unit Emulation
plugins: Add cache miss reporting over a socket.
contrib: Add example hotness monitoring unit server

  include/hw/cxl/cxl.h        |   1 +
  include/hw/cxl/cxl_chmu.h   | 154 ++++++++++++
  include/hw/cxl/cxl_device.h |  13 +-
  include/hw/cxl/cxl_pci.h    |   7 +-
  contrib/hmu/hmu.c           | 312 ++++++++++++++++++++++++
  contrib/plugins/cache.c     |  75 +++++-
  hw/cxl/cxl-chmu.c           | 459 ++++++++++++++++++++++++++++++++++++
  hw/mem/cxl_type3.c          |  25 +-
  hw/cxl/meson.build          |   1 +
  9 files changed, 1035 insertions(+), 12 deletions(-)

create mode 100644 include/hw/cxl/cxl_chmu.h
create mode 100644 contrib/hmu/hmu.c
create mode 100644 hw/cxl/cxl-chmu.c

#7 lore raw
Quoted 23 lines · Jonathan Cameron
Quoted 10 lines · On Fri, 24 Jan 2025 12:55:52 -0800 Pierrick Bouvier…

Hi Jonathan,

thanks for posting this. It's a creative usage of plugins.

I think that your current approach, decoupling plugins, CHMU and device model is a good thing.

I'm not familiar with CXL, but one question that comes to my mind is:
Is that mandatory to do this analysis during execution (vs dumping binary traces from CHMU and plugin and running an analysis post execution)?

Short answer is that post run analysis isn't of much use for developing the OS software story. It works to some degree if you are designing the tracking hardware or algorithms to use that hardware capture a snapshot of hotness - dealing with lack of counters, that sort of thing.

The main intent of this support is to drive live usage of the data in the OS. So it gets this hotness information and migrates more frequently accessed memory to a 'nearer'/lower latency memory node.

Quoted 11 lines

From an OS point of view there will be two ways it uses it:
1) Offline application optimization - that aligns with your suggestion of offline analysis but would typically still need to be live because we have to do the reverse maps and work out what was allocated in particular locations. Not impossible to dump that information from QEMU + the guest OS but the usage flow would then look quite different from what makes sense on real hardware where all the data is available to the host OS directly.
2) Migration of memory. This will dynamically change the PA backing a VA whilst applications are running. The aim being to develop how that happens, we need the dynamic state.

In the end, are you modeling how the real CHMU will work, or simply gathering data to help designing it (number of counters, line size, ...)?

Pierrick

Quoted 198 lines

Jonathan

Quoted 195 lines

Regards,
Pierrick

Quoted 189 lines · Jonathan Cameron

Hi All,

This is an RFC mainly to seek feedback on the approach used, particularly the aspect of how to get data from a TCG plugin into a device model. Two options that we have tried
1. Socket over which the plugin sends data to an external server (as seen here)
2. Register and manage a plugin from within a device model

The external server approach keeps things loosely coupled, but at the cost of separately maintaining that server, protocol definitions etc and some overhead.
The closely couple solution is neater, but I suspect might be controversial (hence I didn't start with that :)

The code here is at best a PoC to illustrate what we have in mind It's not nice code at all, feature gaps, bugs and all! So whilst review is always welcome I'm not requesting it for now.

Kernel support was posted a while back but was done against fake data (still supported here if you don't provide the port parameter to the type3 device)
https://lore.kernel.org/linux-cxl/[email protected]/
I'll post a minor update of that driver shortly to take into account a few specification clarifications but it should work with this without those.

Note there are some other patches on the tree I generated this from so this may not apply to upstream. Easiest is probably to test using gitlab.com/jic23/qemu cxl-2025-01-24

Thanks to Niyas for his suggestions on how to make all this work!

Background
----------

What is the Compute eXpress Link Hotness Monitoring unit and what is it for?
- In a tiered memory equipped server with the slow tier being attached via CXL the expectation is a given workload will benefit from putting data that is frequently fetched from memory in lower latency directly attached DRAM. Less frequently used data can be served from the CXL attached memory with no significant loss of performance. Any data that is hot enough to almost always be in cache doesn't matter as it is only fetch from memory occasionally.
- Working out which memory is best places where is hard to do and in some workloads a dynamic problem. As such we need something we can measure to provide some indication of what data is in the wrong place. There are existing techniques to do this (page faulting, various CPU tracing systems, access bit scanning etc) but they all have significant overheads.
- Monitoring accesses on the CXL device provides a path to getting good data without those overheads. These units are known as CXL Hotness Monitoring Units or CHMUs. Loosely speaking they count accesses to granuals of data (e.g. 4KiB pages). Exactly how they do that and where they sacrifice data accuracy is an implementation trade off.

Why do we need a model that gives real data?
- In general there is a need to develop software on top of these units to move data to the right place. Hard to evaluate that if we are making up the info on what is 'hot'.
- Need to allow for a bunch of 'impdef' solutions. Note that CHMU in this patch set is an oracle - it has enough counters to count every access. That's not realistic but it doesn't get me shouted at by our architecture teams for giving away any secrets. If we move forward with this, I'll probably implement a limited counter + full CAM solution (also unrealistic, but closer to real) I'd be very interested in contributions of other approaches (there are lots in the literature, under the term top-k)
- Resources will be constrained, so whilst a CHMU might in theory allow monitoring everything at once, that will come with a big accuracy cost. We need to design the algorithms that give us good data given those constraints.

So we need a solution to explore the design space and develop the software to take advantage of this hardware (there are various LSF/MM proposals on how to use this an other ways of tracking hotness).
https://lore.kernel.org/all/[email protected]/
https://lore.kernel.org/lkml/Z4XUoWlU-UgRik18@gourry-fedora-PF4VCD3F/T/

QEMU plugins give us a way to do this. In particular the existing Cache plugin can be easily modified to tell use what memory addresses missed at the last level of emulated cache. We can then filter those for the memory address range that maps to CXL and feed them to our counter implementation. On the other side, each instance of CXL type 3 device can connect to this server and request hotness monitoring services + provide parameters etc. Elements such as list threshold management and overflow detection etc are in the CXL HMU QEMU device mode. As noted above, we have an alternative approach that can closely couple things, so the device model registers the plugin directly and there is no server.

How to use it!
--------------

It runs a little slow but it runs and generates somewhat plausible outputs. I'd definitely suggest running it with the pass through optimization patch on the CXL staging tree (and a single direct connected device). Your millage will vary if you try to use other parameters, or hotness units beyond the first one (implementation far from complete!)

To run start the server in contrib/hmu/ providing a port number to listen on.

./chmu 4443

Then launch QEMU with something like the following.

qemu-system-aarch64 -icount shift=1 \
   -plugin ../qemu/bin/native/contrib/plugins/libcache.so,port=4443,missfilterbase=1099511627776,missfiltersize=1099511627776,dcachesize=8192,dassoc=4,dblksize=64,icachesize=8192,iassoc=4,iblksize=64,l2cachesize=32768,l2assoc=16,l2blksize=64 \
   -M virt,ras=on,nvdimm=on,gic-version=3,cxl=on,hmat=on -m 4g,maxmem=8g,slots=4 -cpu max -smp 4 \
   -kernel Image \
   -drive if=none,file=full.qcow2,format=qcow2,id=hd \
   -device pcie-root-port,id=root_port1 \
   -device virtio-blk-pci,drive=hd,x-max-bounce-buffer-size=512k \
   -nographic -no-reboot -append 'earlycon memblock=debug root=/dev/vda2 fsck.mode=skip maxcpus=4 tp_printk' \
   -monitor telnet:127.0.0.1:1234,server,nowait -bios QEMU_EFI.fd \
   -object memory-backend-ram,size=4G,id=mem0 \
   -object memory-backend-file,id=cxl-mem1,share=on,mem-path=/tmp/t3_cxl1.raw,size=1G,align=256M \
   -object memory-backend-file,id=cxl-mem2,share=on,mem-path=/tmp/t3_cxl2.raw,size=1G,align=256M \
   -object memory-backend-file,id=cxl-lsa1,share=on,mem-path=/tmp/t3_lsa1.raw,size=1M,align=1M \
    -object memory-backend-file,id=cxl-mem3,share=on,mem-path=/tmp/t3_cxl3.raw,size=1G,align=256M \
   -object memory-backend-file,id=cxl-mem4,share=on,mem-path=/tmp/t3_cxl4.raw,size=1G,align=256M \
   -object memory-backend-file,id=cxl-lsa2,share=on,mem-path=/tmp/t3_lsa2.raw,size=1M,align=1M \
   -device pxb-cxl,bus_nr=12,bus=pcie.0,id=cxl.1,hdm_for_passthrough=true,numa_node=0\
   -device cxl-rp,port=0,bus=cxl.1,id=cxl_rp_port0,chassis=0,slot=2 \
   -device cxl-type3,bus=cxl_rp_port0,volatile-memdev=cxl-mem1,id=cxl-pmem1,lsa=cxl-lsa1,sn=3,x-speed=32,x-width=16,chmu-port=4443 \
   -machine cxl-fmw.0.targets.0=cxl.1,cxl-fmw.0.size=8G,cxl-fmw.0.interleave-granularity=1k \
   -numa node,nodeid=0,cpus=0-3,memdev=mem0 \
   -numa node,nodeid=1 \
   -object acpi-generic-initiator,id=bob2,pci-dev=bob,node=1 \
   -numa node,nodeid=2 \
   -object acpi-generic-port,id=bob11,pci-bus=cxl.1,node=2 \

In the guest, create and bind the region - this brings up the CXL memory device so accesses go to the memory.

cd /sys/bus/cxl/devices/decoder0.0/
cat create_ram_region
echo region0 > create_ram_region
echo ram > /sys/bus/cxl/devices/decoder2.0/mode echo ram > /sys/bus/cxl/devices/decoder3.0/mode echo $((256 << 21)) > /sys/bus/cxl/devices/decoder2.0/dpa_size cd /sys/bus/cxl/devices/region0/
echo 256 > interleave_granularity
echo 1 > interleave_ways
echo $((256 << 21)) > size
echo decoder2.0 > target0
echo 1 > commit
echo region0 > /sys/bus/cxl/drivers/cxl_region/bind

Finally start perf with something like:

./perf record -a -e cxl_hmu_mem0.0.0/epoch_type=0,access_type=6,\
hotness_threshold=635,epoch_multiplier=4,epoch_scale=4,\
range_base=0,range_size=4096/ ./stress.sh

where stress.sh is

    sleep 2
    numactl --membind 3 stress-ng --vm 1 --vm-bytes 1M --vm-keep -t 5s
    sleep 2

See the results with
./perf report --dump-raw-trace | grep -A 200 HMU

Enjoy and have a good weekend.

Thanks,

Jonathan

Jonathan Cameron (3):
hw/cxl: Initial CXL Hotness Monitoring Unit Emulation
plugins: Add cache miss reporting over a socket.
contrib: Add example hotness monitoring unit server

   include/hw/cxl/cxl.h        |   1 +
   include/hw/cxl/cxl_chmu.h   | 154 ++++++++++++
   include/hw/cxl/cxl_device.h |  13 +-
   include/hw/cxl/cxl_pci.h    |   7 +-
   contrib/hmu/hmu.c           | 312 ++++++++++++++++++++++++
   contrib/plugins/cache.c     |  75 +++++-
   hw/cxl/cxl-chmu.c           | 459 ++++++++++++++++++++++++++++++++++++
   hw/mem/cxl_type3.c          |  25 +-
   hw/cxl/meson.build          |   1 +
   9 files changed, 1035 insertions(+), 12 deletions(-)

create mode 100644 include/hw/cxl/cxl_chmu.h
create mode 100644 contrib/hmu/hmu.c
create mode 100644 hw/cxl/cxl-chmu.c

#8 lore raw
Quoted 36 lines · On Tue, 28 Jan 2025 12:04:19 -0800 Pierrick Bouvier…
Quoted 23 lines · Jonathan Cameron
Quoted 10 lines · On Fri, 24 Jan 2025 12:55:52 -0800 Pierrick Bouvier…

Hi Jonathan,

thanks for posting this. It's a creative usage of plugins.

I think that your current approach, decoupling plugins, CHMU and device model is a good thing.

I'm not familiar with CXL, but one question that comes to my mind is:
Is that mandatory to do this analysis during execution (vs dumping binary traces from CHMU and plugin and running an analysis post execution)?

Short answer is that post run analysis isn't of much use for developing the OS software story. It works to some degree if you are designing the tracking hardware or algorithms to use that hardware capture a snapshot of hotness - dealing with lack of counters, that sort of thing.

The main intent of this support is to drive live usage of the data in the OS. So it gets this hotness information and migrates more frequently accessed memory to a 'nearer'/lower latency memory node.

Quoted 11 lines

From an OS point of view there will be two ways it uses it:
1) Offline application optimization - that aligns with your suggestion of offline analysis but would typically still need to be live because we have to do the reverse maps and work out what was allocated in particular locations. Not impossible to dump that information from QEMU + the guest OS but the usage flow would then look quite different from what makes sense on real hardware where all the data is available to the host OS directly.
2) Migration of memory. This will dynamically change the PA backing a VA whilst applications are running. The aim being to develop how that happens, we need the dynamic state.

Quoted 2 lines

In the end, are you modeling how the real CHMU will work, or simply gathering data to help designing it (number of counters, line size, ...)?

This work is modeling how a real (ish) CHMU will work - particular interest being use in Linux kernel usecases. Otherwise we wouldn't share! :)

For CHMU hardware design, until people reach the live algorithms in the loop stage, tracing techniques and offline analysis tend to be easier to use.

A annoying corner is that the implementations in QEMU will 'probably' remain simplistic because the detailed designs of CHMUs may be considered sensitive. It's a complex space and there are some really interesting and to me surprising approaches.

What we can implement should be good enough for working out the basics of a general software stack but possible it will need tuning against specific implementations. Maybe that necessity will result in more openness on the parts of various uarch / arch teams.

There are some academic works on how to build these trackers, and there should be less sensitivity around those.

This is perhaps an odd corner for QEMU because we are emulating an interface accurately but the hardware behind it intentionally does not have a specification defined implementation and the unusual bit is that implementation affects the output. We can implement a few options that are well defined though.
1) An Oracle ('infinite' counters)
2) Limited counters allocated on first touch in a given epoch (sampling period).

1 is useful for putting an upper bound on data accuracy.
2 is a typical first thing people will look at when considering a implementation.

So conclusion. This is about enabling software development, not tuning a hardware design.

Jonathan

Quoted 202 lines

Pierrick

Quoted 198 lines

Jonathan

Quoted 195 lines

Regards,
Pierrick

Quoted 189 lines · Jonathan Cameron

Hi All,

This is an RFC mainly to seek feedback on the approach used, particularly the aspect of how to get data from a TCG plugin into a device model. Two options that we have tried
1. Socket over which the plugin sends data to an external server (as seen here)
2. Register and manage a plugin from within a device model

The external server approach keeps things loosely coupled, but at the cost of separately maintaining that server, protocol definitions etc and some overhead.
The closely couple solution is neater, but I suspect might be controversial (hence I didn't start with that :)

The code here is at best a PoC to illustrate what we have in mind It's not nice code at all, feature gaps, bugs and all! So whilst review is always welcome I'm not requesting it for now.

Kernel support was posted a while back but was done against fake data (still supported here if you don't provide the port parameter to the type3 device)
https://lore.kernel.org/linux-cxl/[email protected]/
I'll post a minor update of that driver shortly to take into account a few specification clarifications but it should work with this without those.

Note there are some other patches on the tree I generated this from so this may not apply to upstream. Easiest is probably to test using gitlab.com/jic23/qemu cxl-2025-01-24

Thanks to Niyas for his suggestions on how to make all this work!

Background
----------

What is the Compute eXpress Link Hotness Monitoring unit and what is it for?
- In a tiered memory equipped server with the slow tier being attached via CXL the expectation is a given workload will benefit from putting data that is frequently fetched from memory in lower latency directly attached DRAM. Less frequently used data can be served from the CXL attached memory with no significant loss of performance. Any data that is hot enough to almost always be in cache doesn't matter as it is only fetch from memory occasionally.
- Working out which memory is best places where is hard to do and in some workloads a dynamic problem. As such we need something we can measure to provide some indication of what data is in the wrong place. There are existing techniques to do this (page faulting, various CPU tracing systems, access bit scanning etc) but they all have significant overheads.
- Monitoring accesses on the CXL device provides a path to getting good data without those overheads. These units are known as CXL Hotness Monitoring Units or CHMUs. Loosely speaking they count accesses to granuals of data (e.g. 4KiB pages). Exactly how they do that and where they sacrifice data accuracy is an implementation trade off.

Why do we need a model that gives real data?
- In general there is a need to develop software on top of these units to move data to the right place. Hard to evaluate that if we are making up the info on what is 'hot'.
- Need to allow for a bunch of 'impdef' solutions. Note that CHMU in this patch set is an oracle - it has enough counters to count every access. That's not realistic but it doesn't get me shouted at by our architecture teams for giving away any secrets. If we move forward with this, I'll probably implement a limited counter + full CAM solution (also unrealistic, but closer to real) I'd be very interested in contributions of other approaches (there are lots in the literature, under the term top-k)
- Resources will be constrained, so whilst a CHMU might in theory allow monitoring everything at once, that will come with a big accuracy cost. We need to design the algorithms that give us good data given those constraints.

So we need a solution to explore the design space and develop the software to take advantage of this hardware (there are various LSF/MM proposals on how to use this an other ways of tracking hotness).
https://lore.kernel.org/all/[email protected]/
https://lore.kernel.org/lkml/Z4XUoWlU-UgRik18@gourry-fedora-PF4VCD3F/T/

QEMU plugins give us a way to do this. In particular the existing Cache plugin can be easily modified to tell use what memory addresses missed at the last level of emulated cache. We can then filter those for the memory address range that maps to CXL and feed them to our counter implementation. On the other side, each instance of CXL type 3 device can connect to this server and request hotness monitoring services + provide parameters etc. Elements such as list threshold management and overflow detection etc are in the CXL HMU QEMU device mode. As noted above, we have an alternative approach that can closely couple things, so the device model registers the plugin directly and there is no server.

How to use it!
--------------

It runs a little slow but it runs and generates somewhat plausible outputs. I'd definitely suggest running it with the pass through optimization patch on the CXL staging tree (and a single direct connected device). Your millage will vary if you try to use other parameters, or hotness units beyond the first one (implementation far from complete!)

To run start the server in contrib/hmu/ providing a port number to listen on.

./chmu 4443

Then launch QEMU with something like the following.

qemu-system-aarch64 -icount shift=1 \
   -plugin ../qemu/bin/native/contrib/plugins/libcache.so,port=4443,missfilterbase=1099511627776,missfiltersize=1099511627776,dcachesize=8192,dassoc=4,dblksize=64,icachesize=8192,iassoc=4,iblksize=64,l2cachesize=32768,l2assoc=16,l2blksize=64 \
   -M virt,ras=on,nvdimm=on,gic-version=3,cxl=on,hmat=on -m 4g,maxmem=8g,slots=4 -cpu max -smp 4 \
   -kernel Image \
   -drive if=none,file=full.qcow2,format=qcow2,id=hd \
   -device pcie-root-port,id=root_port1 \
   -device virtio-blk-pci,drive=hd,x-max-bounce-buffer-size=512k \
   -nographic -no-reboot -append 'earlycon memblock=debug root=/dev/vda2 fsck.mode=skip maxcpus=4 tp_printk' \
   -monitor telnet:127.0.0.1:1234,server,nowait -bios QEMU_EFI.fd \
   -object memory-backend-ram,size=4G,id=mem0 \
   -object memory-backend-file,id=cxl-mem1,share=on,mem-path=/tmp/t3_cxl1.raw,size=1G,align=256M \
   -object memory-backend-file,id=cxl-mem2,share=on,mem-path=/tmp/t3_cxl2.raw,size=1G,align=256M \
   -object memory-backend-file,id=cxl-lsa1,share=on,mem-path=/tmp/t3_lsa1.raw,size=1M,align=1M \
    -object memory-backend-file,id=cxl-mem3,share=on,mem-path=/tmp/t3_cxl3.raw,size=1G,align=256M \
   -object memory-backend-file,id=cxl-mem4,share=on,mem-path=/tmp/t3_cxl4.raw,size=1G,align=256M \
   -object memory-backend-file,id=cxl-lsa2,share=on,mem-path=/tmp/t3_lsa2.raw,size=1M,align=1M \
   -device pxb-cxl,bus_nr=12,bus=pcie.0,id=cxl.1,hdm_for_passthrough=true,numa_node=0\
   -device cxl-rp,port=0,bus=cxl.1,id=cxl_rp_port0,chassis=0,slot=2 \
   -device cxl-type3,bus=cxl_rp_port0,volatile-memdev=cxl-mem1,id=cxl-pmem1,lsa=cxl-lsa1,sn=3,x-speed=32,x-width=16,chmu-port=4443 \
   -machine cxl-fmw.0.targets.0=cxl.1,cxl-fmw.0.size=8G,cxl-fmw.0.interleave-granularity=1k \
   -numa node,nodeid=0,cpus=0-3,memdev=mem0 \
   -numa node,nodeid=1 \
   -object acpi-generic-initiator,id=bob2,pci-dev=bob,node=1 \
   -numa node,nodeid=2 \
   -object acpi-generic-port,id=bob11,pci-bus=cxl.1,node=2 \

In the guest, create and bind the region - this brings up the CXL memory device so accesses go to the memory.

cd /sys/bus/cxl/devices/decoder0.0/
cat create_ram_region
echo region0 > create_ram_region
echo ram > /sys/bus/cxl/devices/decoder2.0/mode echo ram > /sys/bus/cxl/devices/decoder3.0/mode echo $((256 << 21)) > /sys/bus/cxl/devices/decoder2.0/dpa_size cd /sys/bus/cxl/devices/region0/
echo 256 > interleave_granularity
echo 1 > interleave_ways
echo $((256 << 21)) > size
echo decoder2.0 > target0
echo 1 > commit
echo region0 > /sys/bus/cxl/drivers/cxl_region/bind

Finally start perf with something like:

./perf record -a -e cxl_hmu_mem0.0.0/epoch_type=0,access_type=6,\
hotness_threshold=635,epoch_multiplier=4,epoch_scale=4,\
range_base=0,range_size=4096/ ./stress.sh

where stress.sh is

    sleep 2
    numactl --membind 3 stress-ng --vm 1 --vm-bytes 1M --vm-keep -t 5s
    sleep 2

See the results with
./perf report --dump-raw-trace | grep -A 200 HMU

Enjoy and have a good weekend.

Thanks,

Jonathan

Jonathan Cameron (3):
hw/cxl: Initial CXL Hotness Monitoring Unit Emulation
plugins: Add cache miss reporting over a socket.
contrib: Add example hotness monitoring unit server

   include/hw/cxl/cxl.h        |   1 +
   include/hw/cxl/cxl_chmu.h   | 154 ++++++++++++
   include/hw/cxl/cxl_device.h |  13 +-
   include/hw/cxl/cxl_pci.h    |   7 +-
   contrib/hmu/hmu.c           | 312 ++++++++++++++++++++++++
   contrib/plugins/cache.c     |  75 +++++-
   hw/cxl/cxl-chmu.c           | 459 ++++++++++++++++++++++++++++++++++++
   hw/mem/cxl_type3.c          |  25 +-
   hw/cxl/meson.build          |   1 +
   9 files changed, 1035 insertions(+), 12 deletions(-)

create mode 100644 include/hw/cxl/cxl_chmu.h
create mode 100644 contrib/hmu/hmu.c
create mode 100644 hw/cxl/cxl-chmu.c

#9 lore raw

Hi Jonathan,

Quoted 68 lines · Jonathan Cameron
Quoted 36 lines · On Tue, 28 Jan 2025 12:04:19 -0800 Pierrick Bouvier…
Quoted 23 lines · Jonathan Cameron
Quoted 10 lines · On Fri, 24 Jan 2025 12:55:52 -0800 Pierrick Bouvier…

Hi Jonathan,

thanks for posting this. It's a creative usage of plugins.

I think that your current approach, decoupling plugins, CHMU and device model is a good thing.

I'm not familiar with CXL, but one question that comes to my mind is:
Is that mandatory to do this analysis during execution (vs dumping binary traces from CHMU and plugin and running an analysis post execution)?

Short answer is that post run analysis isn't of much use for developing the OS software story. It works to some degree if you are designing the tracking hardware or algorithms to use that hardware capture a snapshot of hotness - dealing with lack of counters, that sort of thing.

The main intent of this support is to drive live usage of the data in the OS. So it gets this hotness information and migrates more frequently accessed memory to a 'nearer'/lower latency memory node.

Quoted 11 lines

From an OS point of view there will be two ways it uses it:
1) Offline application optimization - that aligns with your suggestion of offline analysis but would typically still need to be live because we have to do the reverse maps and work out what was allocated in particular locations. Not impossible to dump that information from QEMU + the guest OS but the usage flow would then look quite different from what makes sense on real hardware where all the data is available to the host OS directly.
2) Migration of memory. This will dynamically change the PA backing a VA whilst applications are running. The aim being to develop how that happens, we need the dynamic state.

Quoted 2 lines

In the end, are you modeling how the real CHMU will work, or simply gathering data to help designing it (number of counters, line size, ...)?

This work is modeling how a real (ish) CHMU will work - particular interest being use in Linux kernel usecases. Otherwise we wouldn't share! :)

For CHMU hardware design, until people reach the live algorithms in the loop stage, tracing techniques and offline analysis tend to be easier to use.

A annoying corner is that the implementations in QEMU will 'probably' remain simplistic because the detailed designs of CHMUs may be considered sensitive. It's a complex space and there are some really interesting and to me surprising approaches.

What we can implement should be good enough for working out the basics of a general software stack but possible it will need tuning against specific implementations. Maybe that necessity will result in more openness on the parts of various uarch / arch teams.

There are some academic works on how to build these trackers, and there should be less sensitivity around those.

This is perhaps an odd corner for QEMU because we are emulating an interface accurately but the hardware behind it intentionally does not have a specification defined implementation and the unusual bit is that implementation affects the output. We can implement a few options that are well defined though.
1) An Oracle ('infinite' counters)
2) Limited counters allocated on first touch in a given epoch (sampling period).

Quoted 6 lines

1 is useful for putting an upper bound on data accuracy.
2 is a typical first thing people will look at when considering a implementation.

So conclusion. This is about enabling software development, not tuning a hardware design.

Ok, thanks for the clarification.

Considering the approach you followed, as said before, choosing a decoupled solution is the right choice. Plugins should not allow to access internal details of QEMU implementation as a general rule.

Did you think about integrating the server directly in the plugin? From what I understand, the CHMU will contact the server on a per request basis, while instrumentation will contact it for every access.

Beyond communication, the biggest overhead here is to have instrumentation on all memory accesses. Another thing that comes out of my mind is to do a sampling instrumentation. However, it's not easy to do, because the current API does not allow to force a new translation of existing TB, and at translation time, you have no clue whether or not this will be a hot TB. What you can do though, is to ignore most of the accesses, and only send information every 1000 memory accesses. (Note: I have no idea if 1000 is a good threshold :)). It would allow you to skip most of the overhead related to communication and hot pages management.

From what I understood about using plugins, the goal is to track hot pages. Using the cache one is a possibility, but it might be better to use contrib/plugins/hotpages.c instead, or better, a custom simple plugin, simply reporting reads/writes to the server (and implementing sampling as well).

Which slow-down factor (order of magnitude) do you have with this series?

Quoted 205 lines

Jonathan

Quoted 202 lines

Pierrick

Quoted 198 lines

Jonathan

Quoted 195 lines

Regards,
Pierrick

Quoted 189 lines · Jonathan Cameron

Hi All,

This is an RFC mainly to seek feedback on the approach used, particularly the aspect of how to get data from a TCG plugin into a device model. Two options that we have tried
1. Socket over which the plugin sends data to an external server (as seen here)
2. Register and manage a plugin from within a device model

The external server approach keeps things loosely coupled, but at the cost of separately maintaining that server, protocol definitions etc and some overhead.
The closely couple solution is neater, but I suspect might be controversial (hence I didn't start with that :)

The code here is at best a PoC to illustrate what we have in mind It's not nice code at all, feature gaps, bugs and all! So whilst review is always welcome I'm not requesting it for now.

Kernel support was posted a while back but was done against fake data (still supported here if you don't provide the port parameter to the type3 device)
https://lore.kernel.org/linux-cxl/[email protected]/
I'll post a minor update of that driver shortly to take into account a few specification clarifications but it should work with this without those.

Note there are some other patches on the tree I generated this from so this may not apply to upstream. Easiest is probably to test using gitlab.com/jic23/qemu cxl-2025-01-24

Thanks to Niyas for his suggestions on how to make all this work!

Background
----------

What is the Compute eXpress Link Hotness Monitoring unit and what is it for?
- In a tiered memory equipped server with the slow tier being attached via CXL the expectation is a given workload will benefit from putting data that is frequently fetched from memory in lower latency directly attached DRAM. Less frequently used data can be served from the CXL attached memory with no significant loss of performance. Any data that is hot enough to almost always be in cache doesn't matter as it is only fetch from memory occasionally.
- Working out which memory is best places where is hard to do and in some workloads a dynamic problem. As such we need something we can measure to provide some indication of what data is in the wrong place. There are existing techniques to do this (page faulting, various CPU tracing systems, access bit scanning etc) but they all have significant overheads.
- Monitoring accesses on the CXL device provides a path to getting good data without those overheads. These units are known as CXL Hotness Monitoring Units or CHMUs. Loosely speaking they count accesses to granuals of data (e.g. 4KiB pages). Exactly how they do that and where they sacrifice data accuracy is an implementation trade off.

Why do we need a model that gives real data?
- In general there is a need to develop software on top of these units to move data to the right place. Hard to evaluate that if we are making up the info on what is 'hot'.
- Need to allow for a bunch of 'impdef' solutions. Note that CHMU in this patch set is an oracle - it has enough counters to count every access. That's not realistic but it doesn't get me shouted at by our architecture teams for giving away any secrets. If we move forward with this, I'll probably implement a limited counter + full CAM solution (also unrealistic, but closer to real) I'd be very interested in contributions of other approaches (there are lots in the literature, under the term top-k)
- Resources will be constrained, so whilst a CHMU might in theory allow monitoring everything at once, that will come with a big accuracy cost. We need to design the algorithms that give us good data given those constraints.

So we need a solution to explore the design space and develop the software to take advantage of this hardware (there are various LSF/MM proposals on how to use this an other ways of tracking hotness).
https://lore.kernel.org/all/[email protected]/
https://lore.kernel.org/lkml/Z4XUoWlU-UgRik18@gourry-fedora-PF4VCD3F/T/

QEMU plugins give us a way to do this. In particular the existing Cache plugin can be easily modified to tell use what memory addresses missed at the last level of emulated cache. We can then filter those for the memory address range that maps to CXL and feed them to our counter implementation. On the other side, each instance of CXL type 3 device can connect to this server and request hotness monitoring services + provide parameters etc. Elements such as list threshold management and overflow detection etc are in the CXL HMU QEMU device mode. As noted above, we have an alternative approach that can closely couple things, so the device model registers the plugin directly and there is no server.

How to use it!
--------------

It runs a little slow but it runs and generates somewhat plausible outputs. I'd definitely suggest running it with the pass through optimization patch on the CXL staging tree (and a single direct connected device). Your millage will vary if you try to use other parameters, or hotness units beyond the first one (implementation far from complete!)

To run start the server in contrib/hmu/ providing a port number to listen on.

./chmu 4443

Then launch QEMU with something like the following.

qemu-system-aarch64 -icount shift=1 \
    -plugin ../qemu/bin/native/contrib/plugins/libcache.so,port=4443,missfilterbase=1099511627776,missfiltersize=1099511627776,dcachesize=8192,dassoc=4,dblksize=64,icachesize=8192,iassoc=4,iblksize=64,l2cachesize=32768,l2assoc=16,l2blksize=64 \
    -M virt,ras=on,nvdimm=on,gic-version=3,cxl=on,hmat=on -m 4g,maxmem=8g,slots=4 -cpu max -smp 4 \
    -kernel Image \
    -drive if=none,file=full.qcow2,format=qcow2,id=hd \
    -device pcie-root-port,id=root_port1 \
    -device virtio-blk-pci,drive=hd,x-max-bounce-buffer-size=512k \
    -nographic -no-reboot -append 'earlycon memblock=debug root=/dev/vda2 fsck.mode=skip maxcpus=4 tp_printk' \
    -monitor telnet:127.0.0.1:1234,server,nowait -bios QEMU_EFI.fd \
    -object memory-backend-ram,size=4G,id=mem0 \
    -object memory-backend-file,id=cxl-mem1,share=on,mem-path=/tmp/t3_cxl1.raw,size=1G,align=256M \
    -object memory-backend-file,id=cxl-mem2,share=on,mem-path=/tmp/t3_cxl2.raw,size=1G,align=256M \
    -object memory-backend-file,id=cxl-lsa1,share=on,mem-path=/tmp/t3_lsa1.raw,size=1M,align=1M \
     -object memory-backend-file,id=cxl-mem3,share=on,mem-path=/tmp/t3_cxl3.raw,size=1G,align=256M \
    -object memory-backend-file,id=cxl-mem4,share=on,mem-path=/tmp/t3_cxl4.raw,size=1G,align=256M \
    -object memory-backend-file,id=cxl-lsa2,share=on,mem-path=/tmp/t3_lsa2.raw,size=1M,align=1M \
    -device pxb-cxl,bus_nr=12,bus=pcie.0,id=cxl.1,hdm_for_passthrough=true,numa_node=0\
    -device cxl-rp,port=0,bus=cxl.1,id=cxl_rp_port0,chassis=0,slot=2 \
    -device cxl-type3,bus=cxl_rp_port0,volatile-memdev=cxl-mem1,id=cxl-pmem1,lsa=cxl-lsa1,sn=3,x-speed=32,x-width=16,chmu-port=4443 \
    -machine cxl-fmw.0.targets.0=cxl.1,cxl-fmw.0.size=8G,cxl-fmw.0.interleave-granularity=1k \
    -numa node,nodeid=0,cpus=0-3,memdev=mem0 \
    -numa node,nodeid=1 \
    -object acpi-generic-initiator,id=bob2,pci-dev=bob,node=1 \
    -numa node,nodeid=2 \
    -object acpi-generic-port,id=bob11,pci-bus=cxl.1,node=2 \

In the guest, create and bind the region - this brings up the CXL memory device so accesses go to the memory.

cd /sys/bus/cxl/devices/decoder0.0/
cat create_ram_region
echo region0 > create_ram_region
echo ram > /sys/bus/cxl/devices/decoder2.0/mode echo ram > /sys/bus/cxl/devices/decoder3.0/mode echo $((256 << 21)) > /sys/bus/cxl/devices/decoder2.0/dpa_size cd /sys/bus/cxl/devices/region0/
echo 256 > interleave_granularity
echo 1 > interleave_ways
echo $((256 << 21)) > size
echo decoder2.0 > target0
echo 1 > commit
echo region0 > /sys/bus/cxl/drivers/cxl_region/bind

Finally start perf with something like:

./perf record -a -e cxl_hmu_mem0.0.0/epoch_type=0,access_type=6,\
hotness_threshold=635,epoch_multiplier=4,epoch_scale=4,\
range_base=0,range_size=4096/ ./stress.sh

where stress.sh is

     sleep 2
     numactl --membind 3 stress-ng --vm 1 --vm-bytes 1M --vm-keep -t 5s
     sleep 2

See the results with
./perf report --dump-raw-trace | grep -A 200 HMU

Enjoy and have a good weekend.

Thanks,

Jonathan

Jonathan Cameron (3):
hw/cxl: Initial CXL Hotness Monitoring Unit Emulation
plugins: Add cache miss reporting over a socket.
contrib: Add example hotness monitoring unit server

include/hw/cxl/cxl.h | 1 +
include/hw/cxl/cxl_chmu.h | 154 ++++++++++++
include/hw/cxl/cxl_device.h | 13 +-
include/hw/cxl/cxl_pci.h | 7 +-
contrib/hmu/hmu.c | 312 ++++++++++++++++++++++++
contrib/plugins/cache.c | 75 +++++-
hw/cxl/cxl-chmu.c | 459 ++++++++++++++++++++++++++++++++++++
hw/mem/cxl_type3.c | 25 +-
hw/cxl/meson.build | 1 +
9 files changed, 1035 insertions(+), 12 deletions(-) create mode 100644 include/hw/cxl/cxl_chmu.h create mode 100644 contrib/hmu/hmu.c
create mode 100644 hw/cxl/cxl-chmu.c

#10 lore raw
Quoted 80 lines · On Wed, 29 Jan 2025 14:31:10 -0800 Pierrick Bouvier…

Hi Jonathan,

Quoted 68 lines · Jonathan Cameron
Quoted 36 lines · On Tue, 28 Jan 2025 12:04:19 -0800 Pierrick Bouvier…
Quoted 23 lines · Jonathan Cameron
Quoted 10 lines · On Fri, 24 Jan 2025 12:55:52 -0800 Pierrick Bouvier…

Hi Jonathan,

thanks for posting this. It's a creative usage of plugins.

I think that your current approach, decoupling plugins, CHMU and device model is a good thing.

I'm not familiar with CXL, but one question that comes to my mind is:
Is that mandatory to do this analysis during execution (vs dumping binary traces from CHMU and plugin and running an analysis post execution)?

Short answer is that post run analysis isn't of much use for developing the OS software story. It works to some degree if you are designing the tracking hardware or algorithms to use that hardware capture a snapshot of hotness - dealing with lack of counters, that sort of thing.

The main intent of this support is to drive live usage of the data in the OS. So it gets this hotness information and migrates more frequently accessed memory to a 'nearer'/lower latency memory node.

Quoted 11 lines

From an OS point of view there will be two ways it uses it:
1) Offline application optimization - that aligns with your suggestion of offline analysis but would typically still need to be live because we have to do the reverse maps and work out what was allocated in particular locations. Not impossible to dump that information from QEMU + the guest OS but the usage flow would then look quite different from what makes sense on real hardware where all the data is available to the host OS directly.
2) Migration of memory. This will dynamically change the PA backing a VA whilst applications are running. The aim being to develop how that happens, we need the dynamic state.

Quoted 2 lines

In the end, are you modeling how the real CHMU will work, or simply gathering data to help designing it (number of counters, line size, ...)?

This work is modeling how a real (ish) CHMU will work - particular interest being use in Linux kernel usecases. Otherwise we wouldn't share! :)

For CHMU hardware design, until people reach the live algorithms in the loop stage, tracing techniques and offline analysis tend to be easier to use.

A annoying corner is that the implementations in QEMU will 'probably' remain simplistic because the detailed designs of CHMUs may be considered sensitive. It's a complex space and there are some really interesting and to me surprising approaches.

What we can implement should be good enough for working out the basics of a general software stack but possible it will need tuning against specific implementations. Maybe that necessity will result in more openness on the parts of various uarch / arch teams.

There are some academic works on how to build these trackers, and there should be less sensitivity around those.

This is perhaps an odd corner for QEMU because we are emulating an interface accurately but the hardware behind it intentionally does not have a specification defined implementation and the unusual bit is that implementation affects the output. We can implement a few options that are well defined though.
1) An Oracle ('infinite' counters)
2) Limited counters allocated on first touch in a given epoch (sampling period).

Quoted 6 lines

1 is useful for putting an upper bound on data accuracy.
2 is a typical first thing people will look at when considering a implementation.

So conclusion. This is about enabling software development, not tuning a hardware design.

Ok, thanks for the clarification.

Quoted 7 lines

Considering the approach you followed, as said before, choosing a decoupled solution is the right choice. Plugins should not allow to access internal details of QEMU implementation as a general rule.

Did you think about integrating the server directly in the plugin? From what I understand, the CHMU will contact the server on a per request basis, while instrumentation will contact it for every access.

Definitely a possibility. I was a little nervous that would put ordering constraints on plugin load and device creation. I suppose we can connect when the device is enabled by the OS though so that should be easily avoided (if it is a problem at all!). The other thought was that we'd do some of the data handling asynchronously, but that hasn't happened yet and maybe never will + maybe we can do that in a plugin anyway.

The CHMU device model is responsible for timing signals, so it will signal a regular tick (and read back the buffer state to identify overflow etc). So not quite on request but a lot lower data rate than the plugin to CHMU data flow.

Quoted 7 lines

Beyond communication, the biggest overhead here is to have instrumentation on all memory accesses. Another thing that comes out of my mind is to do a sampling instrumentation. However, it's not easy to do, because the current API does not allow to force a new translation of existing TB, and at translation time, you have no clue whether or not this will be a hot TB.

Whilst we will have sampling controls (they are part of the spec, both fixed interval and pseudo random I just haven't wired them up yet), typically sampling and the noise it creates is one of the causes of bad numbers with hot page tracking.

We go to the effort to capture as much data as possible in a short time so that we can move on to focusing the limited tracking resources on another range of the memory. Pushing knowledge of the sampling to the plugin is definitely an option, but we'd still want full resolution hitting the cache simulator as what we want to sample is what misses in the cache.

Quoted 4 lines

What you can do though, is to ignore most of the accesses, and only send information every 1000 memory accesses. (Note: I have no idea if 1000 is a good threshold :)). It would allow you to skip most of the overhead related to communication and hot pages management.

Definitely will implement that as an option when I wire up the rest of the control interface. It can only happen with the knowledge of the OS as it is another thing that will be controlled (or at least taken into account if the device implements only fixed subsampling).

Quoted 6 lines

From what I understood about using plugins, the goal is to track hot
pages. Using the cache one is a possibility, but it might be better to use contrib/plugins/hotpages.c instead, or better, a custom simple plugin, simply reporting reads/writes to the server (and implementing sampling as well).

We need the cache part. Without that the data ends up quite different from what needs to be measured. Data that is in cache and so not fetched much from main memory as a result is (perhaps confusingly) cold. No point in moving it to faster memory and from CXL point of view we never see the accesses anyway so couldn't if we wanted to. Cache simulation is not perfect as we don't have any simulation of prefetchers etc though not sure how much that will hurt us yet. Current cache plugin is obviously simplistic but it is a reasonable starting point.

Also, in some more interesting cases granual it doesn't align with pages (interleaving across multiple devices, much larger regions than pages) so we'd need to switch back to something like this later anyway.

If bringing the hotness monitoring data capture into the plugin it probably makes sense to fork the code to a new plugin anyway as that will add a lot of complexity wherever we add it.

Quoted 2 lines

Which slow-down factor (order of magnitude) do you have with this series?

I've not measured it accurately but not too bad. Maybe 10-20x slow down over TCG (when running arm64 on x86 host). Few minutes to boot a standard distro.

Jonathan

Quoted 208 lines
Quoted 206 lines

Jonathan

Quoted 202 lines

Pierrick

Quoted 198 lines

Jonathan

Quoted 195 lines

Regards,
Pierrick

Quoted 189 lines · Jonathan Cameron

Hi All,

This is an RFC mainly to seek feedback on the approach used, particularly the aspect of how to get data from a TCG plugin into a device model. Two options that we have tried
1. Socket over which the plugin sends data to an external server (as seen here)
2. Register and manage a plugin from within a device model

The external server approach keeps things loosely coupled, but at the cost of separately maintaining that server, protocol definitions etc and some overhead.
The closely couple solution is neater, but I suspect might be controversial (hence I didn't start with that :)

The code here is at best a PoC to illustrate what we have in mind It's not nice code at all, feature gaps, bugs and all! So whilst review is always welcome I'm not requesting it for now.

Kernel support was posted a while back but was done against fake data (still supported here if you don't provide the port parameter to the type3 device)
https://lore.kernel.org/linux-cxl/[email protected]/
I'll post a minor update of that driver shortly to take into account a few specification clarifications but it should work with this without those.

Note there are some other patches on the tree I generated this from so this may not apply to upstream. Easiest is probably to test using gitlab.com/jic23/qemu cxl-2025-01-24

Thanks to Niyas for his suggestions on how to make all this work!

Background
----------

What is the Compute eXpress Link Hotness Monitoring unit and what is it for?
- In a tiered memory equipped server with the slow tier being attached via CXL the expectation is a given workload will benefit from putting data that is frequently fetched from memory in lower latency directly attached DRAM. Less frequently used data can be served from the CXL attached memory with no significant loss of performance. Any data that is hot enough to almost always be in cache doesn't matter as it is only fetch from memory occasionally.
- Working out which memory is best places where is hard to do and in some workloads a dynamic problem. As such we need something we can measure to provide some indication of what data is in the wrong place. There are existing techniques to do this (page faulting, various CPU tracing systems, access bit scanning etc) but they all have significant overheads.
- Monitoring accesses on the CXL device provides a path to getting good data without those overheads. These units are known as CXL Hotness Monitoring Units or CHMUs. Loosely speaking they count accesses to granuals of data (e.g. 4KiB pages). Exactly how they do that and where they sacrifice data accuracy is an implementation trade off.

Why do we need a model that gives real data?
- In general there is a need to develop software on top of these units to move data to the right place. Hard to evaluate that if we are making up the info on what is 'hot'.
- Need to allow for a bunch of 'impdef' solutions. Note that CHMU in this patch set is an oracle - it has enough counters to count every access. That's not realistic but it doesn't get me shouted at by our architecture teams for giving away any secrets. If we move forward with this, I'll probably implement a limited counter + full CAM solution (also unrealistic, but closer to real) I'd be very interested in contributions of other approaches (there are lots in the literature, under the term top-k)
- Resources will be constrained, so whilst a CHMU might in theory allow monitoring everything at once, that will come with a big accuracy cost. We need to design the algorithms that give us good data given those constraints.

So we need a solution to explore the design space and develop the software to take advantage of this hardware (there are various LSF/MM proposals on how to use this an other ways of tracking hotness).
https://lore.kernel.org/all/[email protected]/
https://lore.kernel.org/lkml/Z4XUoWlU-UgRik18@gourry-fedora-PF4VCD3F/T/

QEMU plugins give us a way to do this. In particular the existing Cache plugin can be easily modified to tell use what memory addresses missed at the last level of emulated cache. We can then filter those for the memory address range that maps to CXL and feed them to our counter implementation. On the other side, each instance of CXL type 3 device can connect to this server and request hotness monitoring services + provide parameters etc. Elements such as list threshold management and overflow detection etc are in the CXL HMU QEMU device mode. As noted above, we have an alternative approach that can closely couple things, so the device model registers the plugin directly and there is no server.

How to use it!
--------------

It runs a little slow but it runs and generates somewhat plausible outputs. I'd definitely suggest running it with the pass through optimization patch on the CXL staging tree (and a single direct connected device). Your millage will vary if you try to use other parameters, or hotness units beyond the first one (implementation far from complete!)

To run start the server in contrib/hmu/ providing a port number to listen on.

./chmu 4443

Then launch QEMU with something like the following.

qemu-system-aarch64 -icount shift=1 \
    -plugin ../qemu/bin/native/contrib/plugins/libcache.so,port=4443,missfilterbase=1099511627776,missfiltersize=1099511627776,dcachesize=8192,dassoc=4,dblksize=64,icachesize=8192,iassoc=4,iblksize=64,l2cachesize=32768,l2assoc=16,l2blksize=64 \
    -M virt,ras=on,nvdimm=on,gic-version=3,cxl=on,hmat=on -m 4g,maxmem=8g,slots=4 -cpu max -smp 4 \
    -kernel Image \
    -drive if=none,file=full.qcow2,format=qcow2,id=hd \
    -device pcie-root-port,id=root_port1 \
    -device virtio-blk-pci,drive=hd,x-max-bounce-buffer-size=512k \
    -nographic -no-reboot -append 'earlycon memblock=debug root=/dev/vda2 fsck.mode=skip maxcpus=4 tp_printk' \
    -monitor telnet:127.0.0.1:1234,server,nowait -bios QEMU_EFI.fd \
    -object memory-backend-ram,size=4G,id=mem0 \
    -object memory-backend-file,id=cxl-mem1,share=on,mem-path=/tmp/t3_cxl1.raw,size=1G,align=256M \
    -object memory-backend-file,id=cxl-mem2,share=on,mem-path=/tmp/t3_cxl2.raw,size=1G,align=256M \
    -object memory-backend-file,id=cxl-lsa1,share=on,mem-path=/tmp/t3_lsa1.raw,size=1M,align=1M \
     -object memory-backend-file,id=cxl-mem3,share=on,mem-path=/tmp/t3_cxl3.raw,size=1G,align=256M \
    -object memory-backend-file,id=cxl-mem4,share=on,mem-path=/tmp/t3_cxl4.raw,size=1G,align=256M \
    -object memory-backend-file,id=cxl-lsa2,share=on,mem-path=/tmp/t3_lsa2.raw,size=1M,align=1M \
    -device pxb-cxl,bus_nr=12,bus=pcie.0,id=cxl.1,hdm_for_passthrough=true,numa_node=0\
    -device cxl-rp,port=0,bus=cxl.1,id=cxl_rp_port0,chassis=0,slot=2 \
    -device cxl-type3,bus=cxl_rp_port0,volatile-memdev=cxl-mem1,id=cxl-pmem1,lsa=cxl-lsa1,sn=3,x-speed=32,x-width=16,chmu-port=4443 \
    -machine cxl-fmw.0.targets.0=cxl.1,cxl-fmw.0.size=8G,cxl-fmw.0.interleave-granularity=1k \
    -numa node,nodeid=0,cpus=0-3,memdev=mem0 \
    -numa node,nodeid=1 \
    -object acpi-generic-initiator,id=bob2,pci-dev=bob,node=1 \
    -numa node,nodeid=2 \
    -object acpi-generic-port,id=bob11,pci-bus=cxl.1,node=2 \

In the guest, create and bind the region - this brings up the CXL memory device so accesses go to the memory.

cd /sys/bus/cxl/devices/decoder0.0/
cat create_ram_region
echo region0 > create_ram_region
echo ram > /sys/bus/cxl/devices/decoder2.0/mode echo ram > /sys/bus/cxl/devices/decoder3.0/mode echo $((256 << 21)) > /sys/bus/cxl/devices/decoder2.0/dpa_size cd /sys/bus/cxl/devices/region0/
echo 256 > interleave_granularity
echo 1 > interleave_ways
echo $((256 << 21)) > size
echo decoder2.0 > target0
echo 1 > commit
echo region0 > /sys/bus/cxl/drivers/cxl_region/bind

Finally start perf with something like:

./perf record -a -e cxl_hmu_mem0.0.0/epoch_type=0,access_type=6,\
hotness_threshold=635,epoch_multiplier=4,epoch_scale=4,\
range_base=0,range_size=4096/ ./stress.sh

where stress.sh is

     sleep 2
     numactl --membind 3 stress-ng --vm 1 --vm-bytes 1M --vm-keep -t 5s
     sleep 2

See the results with
./perf report --dump-raw-trace | grep -A 200 HMU

Enjoy and have a good weekend.

Thanks,

Jonathan

Jonathan Cameron (3):
hw/cxl: Initial CXL Hotness Monitoring Unit Emulation
plugins: Add cache miss reporting over a socket.
contrib: Add example hotness monitoring unit server

include/hw/cxl/cxl.h | 1 +
include/hw/cxl/cxl_chmu.h | 154 ++++++++++++
include/hw/cxl/cxl_device.h | 13 +-
include/hw/cxl/cxl_pci.h | 7 +-
contrib/hmu/hmu.c | 312 ++++++++++++++++++++++++
contrib/plugins/cache.c | 75 +++++-
hw/cxl/cxl-chmu.c | 459 ++++++++++++++++++++++++++++++++++++
hw/mem/cxl_type3.c | 25 +-
hw/cxl/meson.build | 1 +
9 files changed, 1035 insertions(+), 12 deletions(-) create mode 100644 include/hw/cxl/cxl_chmu.h create mode 100644 contrib/hmu/hmu.c
create mode 100644 hw/cxl/cxl-chmu.c

#11 lore raw
Quoted 91 lines · Jonathan Cameron
Quoted 80 lines · On Wed, 29 Jan 2025 14:31:10 -0800 Pierrick Bouvier…

Hi Jonathan,

Quoted 68 lines · Jonathan Cameron
Quoted 36 lines · On Tue, 28 Jan 2025 12:04:19 -0800 Pierrick Bouvier…
Quoted 23 lines · Jonathan Cameron
Quoted 10 lines · On Fri, 24 Jan 2025 12:55:52 -0800 Pierrick Bouvier…

Hi Jonathan,

thanks for posting this. It's a creative usage of plugins.

I think that your current approach, decoupling plugins, CHMU and device model is a good thing.

I'm not familiar with CXL, but one question that comes to my mind is:
Is that mandatory to do this analysis during execution (vs dumping binary traces from CHMU and plugin and running an analysis post execution)?

Short answer is that post run analysis isn't of much use for developing the OS software story. It works to some degree if you are designing the tracking hardware or algorithms to use that hardware capture a snapshot of hotness - dealing with lack of counters, that sort of thing.

The main intent of this support is to drive live usage of the data in the OS. So it gets this hotness information and migrates more frequently accessed memory to a 'nearer'/lower latency memory node.

Quoted 11 lines

From an OS point of view there will be two ways it uses it:
1) Offline application optimization - that aligns with your suggestion of offline analysis but would typically still need to be live because we have to do the reverse maps and work out what was allocated in particular locations. Not impossible to dump that information from QEMU + the guest OS but the usage flow would then look quite different from what makes sense on real hardware where all the data is available to the host OS directly.
2) Migration of memory. This will dynamically change the PA backing a VA whilst applications are running. The aim being to develop how that happens, we need the dynamic state.

Quoted 2 lines

In the end, are you modeling how the real CHMU will work, or simply gathering data to help designing it (number of counters, line size, ...)?

This work is modeling how a real (ish) CHMU will work - particular interest being use in Linux kernel usecases. Otherwise we wouldn't share! :)

For CHMU hardware design, until people reach the live algorithms in the loop stage, tracing techniques and offline analysis tend to be easier to use.

A annoying corner is that the implementations in QEMU will 'probably' remain simplistic because the detailed designs of CHMUs may be considered sensitive. It's a complex space and there are some really interesting and to me surprising approaches.

What we can implement should be good enough for working out the basics of a general software stack but possible it will need tuning against specific implementations. Maybe that necessity will result in more openness on the parts of various uarch / arch teams.

There are some academic works on how to build these trackers, and there should be less sensitivity around those.

This is perhaps an odd corner for QEMU because we are emulating an interface accurately but the hardware behind it intentionally does not have a specification defined implementation and the unusual bit is that implementation affects the output. We can implement a few options that are well defined though.
1) An Oracle ('infinite' counters)
2) Limited counters allocated on first touch in a given epoch (sampling period).

Quoted 6 lines

1 is useful for putting an upper bound on data accuracy.
2 is a typical first thing people will look at when considering a implementation.

So conclusion. This is about enabling software development, not tuning a hardware design.

Ok, thanks for the clarification.

Quoted 7 lines

Considering the approach you followed, as said before, choosing a decoupled solution is the right choice. Plugins should not allow to access internal details of QEMU implementation as a general rule.

Did you think about integrating the server directly in the plugin? From what I understand, the CHMU will contact the server on a per request basis, while instrumentation will contact it for every access.

Quoted 7 lines

Definitely a possibility. I was a little nervous that would put ordering constraints on plugin load and device creation. I suppose we can connect when the device is enabled by the OS though so that should be easily avoided (if it is a problem at all!). The other thought was that we'd do some of the data handling asynchronously, but that hasn't happened yet and maybe never will + maybe we can do that in a plugin anyway.

You still need a separate thread triggered from the plugin install function for the server, and initiating the connection first time you need it seems like a good approach.

However, in the plugin itself, you now have to deal with concurrency issues, since several cpus can report memory accesses at the same time, which were sequentialized through the socket before.

Quoted 18 lines

The CHMU device model is responsible for timing signals, so it will signal a regular tick (and read back the buffer state to identify overflow etc). So not quite on request but a lot lower data rate than the plugin to CHMU data flow.

Quoted 7 lines

Beyond communication, the biggest overhead here is to have instrumentation on all memory accesses. Another thing that comes out of my mind is to do a sampling instrumentation. However, it's not easy to do, because the current API does not allow to force a new translation of existing TB, and at translation time, you have no clue whether or not this will be a hot TB.

Whilst we will have sampling controls (they are part of the spec, both fixed interval and pseudo random I just haven't wired them up yet), typically sampling and the noise it creates is one of the causes of bad numbers with hot page tracking.

Quoted 6 lines

We go to the effort to capture as much data as possible in a short time so that we can move on to focusing the limited tracking resources on another range of the memory. Pushing knowledge of the sampling to the plugin is definitely an option, but we'd still want full resolution hitting the cache simulator as what we want to sample is what misses in the cache.

Sure, I was just proposing this to speed up execution. There is a trade-off concerning accuracy. However, I'm surprised you observed such bad numbers (no data locality in your use cases?)

Quoted 27 lines
Quoted 4 lines

What you can do though, is to ignore most of the accesses, and only send information every 1000 memory accesses. (Note: I have no idea if 1000 is a good threshold :)). It would allow you to skip most of the overhead related to communication and hot pages management.

Definitely will implement that as an option when I wire up the rest of the control interface. It can only happen with the knowledge of the OS as it is another thing that will be controlled (or at least taken into account if the device implements only fixed subsampling).

Quoted 6 lines

From what I understood about using plugins, the goal is to track hot
pages. Using the cache one is a possibility, but it might be better to use contrib/plugins/hotpages.c instead, or better, a custom simple plugin, simply reporting reads/writes to the server (and implementing sampling as well).

We need the cache part. Without that the data ends up quite different from what needs to be measured. Data that is in cache and so not fetched much from main memory as a result is (perhaps confusingly) cold. No point in moving it to faster memory and from CXL point of view we never see the accesses anyway so couldn't if we wanted to. Cache simulation is not perfect as we don't have any simulation of prefetchers etc though not sure how much that will hurt us yet. Current cache plugin is obviously simplistic but it is a reasonable starting point.

Quoted 8 lines

Also, in some more interesting cases granual it doesn't align with pages (interleaving across multiple devices, much larger regions than pages) so we'd need to switch back to something like this later anyway.

If bringing the hotness monitoring data capture into the plugin it probably makes sense to fork the code to a new plugin anyway as that will add a lot of complexity wherever we add it.

From what I understand, the goal is to keep hot pages in a specific CXL cache, located between the cpu cache, and DRAM, and to avoid keeping the same data than in cpu cache. However, cache topology will be different for various cpus, with different levels, different sizes, and different associativity.
Thus, it seems that it will be impossible to model anything, simply based on a generic "cache" algorithm.

The thing that puzzles me in the approach is that we compare a cpu data cache (based on small chunks of data), and a hot page cache (based on a page size). It seems normal to hold the same data, and be able to provide the rest of the page to cpu cache when needed.

Quoted 217 lines
Quoted 2 lines

Which slow-down factor (order of magnitude) do you have with this series?

I've not measured it accurately but not too bad. Maybe 10-20x slow down over TCG (when running arm64 on x86 host). Few minutes to boot a standard distro.

Jonathan

Quoted 208 lines
Quoted 206 lines

Jonathan

Quoted 202 lines

Pierrick

Quoted 198 lines

Jonathan

Quoted 195 lines

Regards,
Pierrick

Quoted 189 lines · Jonathan Cameron

Hi All,

This is an RFC mainly to seek feedback on the approach used, particularly the aspect of how to get data from a TCG plugin into a device model. Two options that we have tried
1. Socket over which the plugin sends data to an external server (as seen here)
2. Register and manage a plugin from within a device model

The external server approach keeps things loosely coupled, but at the cost of separately maintaining that server, protocol definitions etc and some overhead.
The closely couple solution is neater, but I suspect might be controversial (hence I didn't start with that :)

The code here is at best a PoC to illustrate what we have in mind It's not nice code at all, feature gaps, bugs and all! So whilst review is always welcome I'm not requesting it for now.

Kernel support was posted a while back but was done against fake data (still supported here if you don't provide the port parameter to the type3 device)
https://lore.kernel.org/linux-cxl/[email protected]/
I'll post a minor update of that driver shortly to take into account a few specification clarifications but it should work with this without those.

Note there are some other patches on the tree I generated this from so this may not apply to upstream. Easiest is probably to test using gitlab.com/jic23/qemu cxl-2025-01-24

Thanks to Niyas for his suggestions on how to make all this work!

Background
----------

What is the Compute eXpress Link Hotness Monitoring unit and what is it for?
- In a tiered memory equipped server with the slow tier being attached via CXL the expectation is a given workload will benefit from putting data that is frequently fetched from memory in lower latency directly attached DRAM. Less frequently used data can be served from the CXL attached memory with no significant loss of performance. Any data that is hot enough to almost always be in cache doesn't matter as it is only fetch from memory occasionally.
- Working out which memory is best places where is hard to do and in some workloads a dynamic problem. As such we need something we can measure to provide some indication of what data is in the wrong place. There are existing techniques to do this (page faulting, various CPU tracing systems, access bit scanning etc) but they all have significant overheads.
- Monitoring accesses on the CXL device provides a path to getting good data without those overheads. These units are known as CXL Hotness Monitoring Units or CHMUs. Loosely speaking they count accesses to granuals of data (e.g. 4KiB pages). Exactly how they do that and where they sacrifice data accuracy is an implementation trade off.

Why do we need a model that gives real data?
- In general there is a need to develop software on top of these units to move data to the right place. Hard to evaluate that if we are making up the info on what is 'hot'.
- Need to allow for a bunch of 'impdef' solutions. Note that CHMU in this patch set is an oracle - it has enough counters to count every access. That's not realistic but it doesn't get me shouted at by our architecture teams for giving away any secrets. If we move forward with this, I'll probably implement a limited counter + full CAM solution (also unrealistic, but closer to real) I'd be very interested in contributions of other approaches (there are lots in the literature, under the term top-k)
- Resources will be constrained, so whilst a CHMU might in theory allow monitoring everything at once, that will come with a big accuracy cost. We need to design the algorithms that give us good data given those constraints.

So we need a solution to explore the design space and develop the software to take advantage of this hardware (there are various LSF/MM proposals on how to use this an other ways of tracking hotness).
https://lore.kernel.org/all/[email protected]/
https://lore.kernel.org/lkml/Z4XUoWlU-UgRik18@gourry-fedora-PF4VCD3F/T/

QEMU plugins give us a way to do this. In particular the existing Cache plugin can be easily modified to tell use what memory addresses missed at the last level of emulated cache. We can then filter those for the memory address range that maps to CXL and feed them to our counter implementation. On the other side, each instance of CXL type 3 device can connect to this server and request hotness monitoring services + provide parameters etc. Elements such as list threshold management and overflow detection etc are in the CXL HMU QEMU device mode. As noted above, we have an alternative approach that can closely couple things, so the device model registers the plugin directly and there is no server.

How to use it!
--------------

It runs a little slow but it runs and generates somewhat plausible outputs. I'd definitely suggest running it with the pass through optimization patch on the CXL staging tree (and a single direct connected device). Your millage will vary if you try to use other parameters, or hotness units beyond the first one (implementation far from complete!)

To run start the server in contrib/hmu/ providing a port number to listen on.

./chmu 4443

Then launch QEMU with something like the following.

qemu-system-aarch64 -icount shift=1 \
     -plugin ../qemu/bin/native/contrib/plugins/libcache.so,port=4443,missfilterbase=1099511627776,missfiltersize=1099511627776,dcachesize=8192,dassoc=4,dblksize=64,icachesize=8192,iassoc=4,iblksize=64,l2cachesize=32768,l2assoc=16,l2blksize=64 \
     -M virt,ras=on,nvdimm=on,gic-version=3,cxl=on,hmat=on -m 4g,maxmem=8g,slots=4 -cpu max -smp 4 \
     -kernel Image \
     -drive if=none,file=full.qcow2,format=qcow2,id=hd \
     -device pcie-root-port,id=root_port1 \
     -device virtio-blk-pci,drive=hd,x-max-bounce-buffer-size=512k \
     -nographic -no-reboot -append 'earlycon memblock=debug root=/dev/vda2 fsck.mode=skip maxcpus=4 tp_printk' \
     -monitor telnet:127.0.0.1:1234,server,nowait -bios QEMU_EFI.fd \
     -object memory-backend-ram,size=4G,id=mem0 \
     -object memory-backend-file,id=cxl-mem1,share=on,mem-path=/tmp/t3_cxl1.raw,size=1G,align=256M \
     -object memory-backend-file,id=cxl-mem2,share=on,mem-path=/tmp/t3_cxl2.raw,size=1G,align=256M \
     -object memory-backend-file,id=cxl-lsa1,share=on,mem-path=/tmp/t3_lsa1.raw,size=1M,align=1M \
      -object memory-backend-file,id=cxl-mem3,share=on,mem-path=/tmp/t3_cxl3.raw,size=1G,align=256M \
     -object memory-backend-file,id=cxl-mem4,share=on,mem-path=/tmp/t3_cxl4.raw,size=1G,align=256M \
     -object memory-backend-file,id=cxl-lsa2,share=on,mem-path=/tmp/t3_lsa2.raw,size=1M,align=1M \
     -device pxb-cxl,bus_nr=12,bus=pcie.0,id=cxl.1,hdm_for_passthrough=true,numa_node=0\
     -device cxl-rp,port=0,bus=cxl.1,id=cxl_rp_port0,chassis=0,slot=2 \
     -device cxl-type3,bus=cxl_rp_port0,volatile-memdev=cxl-mem1,id=cxl-pmem1,lsa=cxl-lsa1,sn=3,x-speed=32,x-width=16,chmu-port=4443 \
     -machine cxl-fmw.0.targets.0=cxl.1,cxl-fmw.0.size=8G,cxl-fmw.0.interleave-granularity=1k \
     -numa node,nodeid=0,cpus=0-3,memdev=mem0 \
     -numa node,nodeid=1 \
     -object acpi-generic-initiator,id=bob2,pci-dev=bob,node=1 \
     -numa node,nodeid=2 \
     -object acpi-generic-port,id=bob11,pci-bus=cxl.1,node=2 \

In the guest, create and bind the region - this brings up the CXL memory device so accesses go to the memory.

cd /sys/bus/cxl/devices/decoder0.0/
cat create_ram_region
echo region0 > create_ram_region
echo ram > /sys/bus/cxl/devices/decoder2.0/mode echo ram > /sys/bus/cxl/devices/decoder3.0/mode echo $((256 << 21)) > /sys/bus/cxl/devices/decoder2.0/dpa_size cd /sys/bus/cxl/devices/region0/
echo 256 > interleave_granularity
echo 1 > interleave_ways
echo $((256 << 21)) > size
echo decoder2.0 > target0
echo 1 > commit
echo region0 > /sys/bus/cxl/drivers/cxl_region/bind

Finally start perf with something like:

./perf record -a -e cxl_hmu_mem0.0.0/epoch_type=0,access_type=6,\
hotness_threshold=635,epoch_multiplier=4,epoch_scale=4,\
range_base=0,range_size=4096/ ./stress.sh

where stress.sh is

      sleep 2
      numactl --membind 3 stress-ng --vm 1 --vm-bytes 1M --vm-keep -t 5s
      sleep 2

See the results with
./perf report --dump-raw-trace | grep -A 200 HMU

Enjoy and have a good weekend.

Thanks,

Jonathan

Jonathan Cameron (3):
hw/cxl: Initial CXL Hotness Monitoring Unit Emulation
plugins: Add cache miss reporting over a socket.
contrib: Add example hotness monitoring unit server

include/hw/cxl/cxl.h | 1 +
include/hw/cxl/cxl_chmu.h | 154 ++++++++++++
include/hw/cxl/cxl_device.h | 13 +-
include/hw/cxl/cxl_pci.h | 7 +-
contrib/hmu/hmu.c | 312 ++++++++++++++++++++++++
contrib/plugins/cache.c | 75 +++++-
hw/cxl/cxl-chmu.c | 459 ++++++++++++++++++++++++++++++++++++
hw/mem/cxl_type3.c | 25 +-
hw/cxl/meson.build | 1 +
9 files changed, 1035 insertions(+), 12 deletions(-) create mode 100644 include/hw/cxl/cxl_chmu.h create mode 100644 contrib/hmu/hmu.c
create mode 100644 hw/cxl/cxl-chmu.c

#12 lore raw
Quoted 100 lines · On Thu, 30 Jan 2025 10:28:03 -0800 Pierrick Bouvier…
Quoted 91 lines · Jonathan Cameron
Quoted 80 lines · On Wed, 29 Jan 2025 14:31:10 -0800 Pierrick Bouvier…

Hi Jonathan,

Quoted 68 lines · Jonathan Cameron
Quoted 36 lines · On Tue, 28 Jan 2025 12:04:19 -0800 Pierrick Bouvier…
Quoted 23 lines · Jonathan Cameron
Quoted 10 lines · On Fri, 24 Jan 2025 12:55:52 -0800 Pierrick Bouvier…

Hi Jonathan,

thanks for posting this. It's a creative usage of plugins.

I think that your current approach, decoupling plugins, CHMU and device model is a good thing.

I'm not familiar with CXL, but one question that comes to my mind is:
Is that mandatory to do this analysis during execution (vs dumping binary traces from CHMU and plugin and running an analysis post execution)?

Short answer is that post run analysis isn't of much use for developing the OS software story. It works to some degree if you are designing the tracking hardware or algorithms to use that hardware capture a snapshot of hotness - dealing with lack of counters, that sort of thing.

The main intent of this support is to drive live usage of the data in the OS. So it gets this hotness information and migrates more frequently accessed memory to a 'nearer'/lower latency memory node.

Quoted 11 lines

From an OS point of view there will be two ways it uses it:
1) Offline application optimization - that aligns with your suggestion of offline analysis but would typically still need to be live because we have to do the reverse maps and work out what was allocated in particular locations. Not impossible to dump that information from QEMU + the guest OS but the usage flow would then look quite different from what makes sense on real hardware where all the data is available to the host OS directly.
2) Migration of memory. This will dynamically change the PA backing a VA whilst applications are running. The aim being to develop how that happens, we need the dynamic state.

Quoted 2 lines

In the end, are you modeling how the real CHMU will work, or simply gathering data to help designing it (number of counters, line size, ...)?

This work is modeling how a real (ish) CHMU will work - particular interest being use in Linux kernel usecases. Otherwise we wouldn't share! :)

For CHMU hardware design, until people reach the live algorithms in the loop stage, tracing techniques and offline analysis tend to be easier to use.

A annoying corner is that the implementations in QEMU will 'probably' remain simplistic because the detailed designs of CHMUs may be considered sensitive. It's a complex space and there are some really interesting and to me surprising approaches.

What we can implement should be good enough for working out the basics of a general software stack but possible it will need tuning against specific implementations. Maybe that necessity will result in more openness on the parts of various uarch / arch teams.

There are some academic works on how to build these trackers, and there should be less sensitivity around those.

This is perhaps an odd corner for QEMU because we are emulating an interface accurately but the hardware behind it intentionally does not have a specification defined implementation and the unusual bit is that implementation affects the output. We can implement a few options that are well defined though.
1) An Oracle ('infinite' counters)
2) Limited counters allocated on first touch in a given epoch (sampling period).

Quoted 6 lines

1 is useful for putting an upper bound on data accuracy.
2 is a typical first thing people will look at when considering a implementation.

So conclusion. This is about enabling software development, not tuning a hardware design.

Ok, thanks for the clarification.

Quoted 7 lines

Considering the approach you followed, as said before, choosing a decoupled solution is the right choice. Plugins should not allow to access internal details of QEMU implementation as a general rule.

Did you think about integrating the server directly in the plugin? From what I understand, the CHMU will contact the server on a per request basis, while instrumentation will contact it for every access.

Quoted 7 lines

Definitely a possibility. I was a little nervous that would put ordering constraints on plugin load and device creation. I suppose we can connect when the device is enabled by the OS though so that should be easily avoided (if it is a problem at all!). The other thought was that we'd do some of the data handling asynchronously, but that hasn't happened yet and maybe never will + maybe we can do that in a plugin anyway.

Quoted 7 lines

You still need a separate thread triggered from the plugin install function for the server, and initiating the connection first time you need it seems like a good approach.

However, in the plugin itself, you now have to deal with concurrency issues, since several cpus can report memory accesses at the same time, which were sequentialized through the socket before.

True. There is already serialization going on for shared l2 caches but needs something a little more general.

Quoted 26 lines
Quoted 19 lines

The CHMU device model is responsible for timing signals, so it will signal a regular tick (and read back the buffer state to identify overflow etc). So not quite on request but a lot lower data rate than the plugin to CHMU data flow.

Quoted 7 lines

Beyond communication, the biggest overhead here is to have instrumentation on all memory accesses. Another thing that comes out of my mind is to do a sampling instrumentation. However, it's not easy to do, because the current API does not allow to force a new translation of existing TB, and at translation time, you have no clue whether or not this will be a hot TB.

Whilst we will have sampling controls (they are part of the spec, both fixed interval and pseudo random I just haven't wired them up yet), typically sampling and the noise it creates is one of the causes of bad numbers with hot page tracking.

Quoted 6 lines

We go to the effort to capture as much data as possible in a short time so that we can move on to focusing the limited tracking resources on another range of the memory. Pushing knowledge of the sampling to the plugin is definitely an option, but we'd still want full resolution hitting the cache simulator as what we want to sample is what misses in the cache.

Quoted 3 lines

Sure, I was just proposing this to speed up execution. There is a trade-off concerning accuracy. However, I'm surprised you observed such bad numbers (no data locality in your use cases?)

Not at all - in fact kind of the opposite. Mostly I've just seen how bad the data is and not looked too closely at why.

I think what can happen is a that a clump of accesses to a page come together as streaming across the cachelines of a page.

If you subsample the data you hit maybe one or two of those rather than a much higher count. Have that happen a few times and your subsampled data badly missrepresents the actual accesses. Arguably those streaming cases might not matter to performance (prefetcher heaven) but they effect what we see when trying to match what the hardware does.
So unless the hardware has been specifically told to subsample we should not do so.

Quoted 32 lines
Quoted 32 lines
Quoted 5 lines

What you can do though, is to ignore most of the accesses, and only send information every 1000 memory accesses. (Note: I have no idea if 1000 is a good threshold :)). It would allow you to skip most of the overhead related to communication and hot pages management.

Definitely will implement that as an option when I wire up the rest of the control interface. It can only happen with the knowledge of the OS as it is another thing that will be controlled (or at least taken into account if the device implements only fixed subsampling).

Quoted 6 lines

From what I understood about using plugins, the goal is to track hot
pages. Using the cache one is a possibility, but it might be better to use contrib/plugins/hotpages.c instead, or better, a custom simple plugin, simply reporting reads/writes to the server (and implementing sampling as well).

We need the cache part. Without that the data ends up quite different from what needs to be measured. Data that is in cache and so not fetched much from main memory as a result is (perhaps confusingly) cold. No point in moving it to faster memory and from CXL point of view we never see the accesses anyway so couldn't if we wanted to. Cache simulation is not perfect as we don't have any simulation of prefetchers etc though not sure how much that will hurt us yet. Current cache plugin is obviously simplistic but it is a reasonable starting point.

Also, in some more interesting cases granual it doesn't align with pages (interleaving across multiple devices, much larger regions than pages) so we'd need to switch back to something like this later anyway.

Quoted 8 lines
Quoted 4 lines

If bringing the hotness monitoring data capture into the plugin it probably makes sense to fork the code to a new plugin anyway as that will add a lot of complexity wherever we add it.

From what I understand, the goal is to keep hot pages in a specific CXL
cache, located between the cpu cache, and DRAM, and to avoid keeping the same data than in cpu cache.

No. This is about tiering across CXL and other memory in the system with software migrating that memory (similar to NUMA balancing). Typically your low latency memory that you are migrating to is direct attached DDR4/5. The high latency tier is the CXL memory. There are no extra caches in this set up this is targetting.

Quoted 4 lines

However, cache topology will be different for various cpus, with different levels, different sizes, and different associativity. Thus, it seems that it will be impossible to model anything, simply based on a generic "cache" algorithm.

Sure we may need to improve that cache part of the plugin. Right now it is indeed simplistic. However, aim here is to get 'a reasonable' software stack up. Emulation won't be perfect but at least the numbers should be somewhat close to the real ones. Obviously anyone doing system design should have a suitable cycle accurate model, but that's useless for the software stack design (too slow by far!!!)

Quoted 5 lines

The thing that puzzles me in the approach is that we compare a cpu data cache (based on small chunks of data), and a hot page cache (based on a page size). It seems normal to hold the same data, and be able to provide the rest of the page to cpu cache when needed.

This is the confusion above. This is not in any way about a special cache. There are other things you can do in that space, but this is just fast and slow dram + the normal CPU caching heirarchy with the OS choosing to migrate the pages (and map the VA to PA appropriately to point to the new PA).

Jonathan

Quoted 219 lines
Quoted 218 lines
Quoted 3 lines

Which slow-down factor (order of magnitude) do you have with this series?

I've not measured it accurately but not too bad. Maybe 10-20x slow down over TCG (when running arm64 on x86 host). Few minutes to boot a standard distro.

Jonathan

Quoted 208 lines
Quoted 206 lines

Jonathan

Quoted 202 lines

Pierrick

Quoted 198 lines

Jonathan

Quoted 195 lines

Regards,
Pierrick

Quoted 189 lines · Jonathan Cameron

Hi All,

This is an RFC mainly to seek feedback on the approach used, particularly the aspect of how to get data from a TCG plugin into a device model. Two options that we have tried
1. Socket over which the plugin sends data to an external server (as seen here)
2. Register and manage a plugin from within a device model

The external server approach keeps things loosely coupled, but at the cost of separately maintaining that server, protocol definitions etc and some overhead.
The closely couple solution is neater, but I suspect might be controversial (hence I didn't start with that :)

The code here is at best a PoC to illustrate what we have in mind It's not nice code at all, feature gaps, bugs and all! So whilst review is always welcome I'm not requesting it for now.

Kernel support was posted a while back but was done against fake data (still supported here if you don't provide the port parameter to the type3 device)
https://lore.kernel.org/linux-cxl/[email protected]/
I'll post a minor update of that driver shortly to take into account a few specification clarifications but it should work with this without those.

Note there are some other patches on the tree I generated this from so this may not apply to upstream. Easiest is probably to test using gitlab.com/jic23/qemu cxl-2025-01-24

Thanks to Niyas for his suggestions on how to make all this work!

Background
----------

What is the Compute eXpress Link Hotness Monitoring unit and what is it for?
- In a tiered memory equipped server with the slow tier being attached via CXL the expectation is a given workload will benefit from putting data that is frequently fetched from memory in lower latency directly attached DRAM. Less frequently used data can be served from the CXL attached memory with no significant loss of performance. Any data that is hot enough to almost always be in cache doesn't matter as it is only fetch from memory occasionally.
- Working out which memory is best places where is hard to do and in some workloads a dynamic problem. As such we need something we can measure to provide some indication of what data is in the wrong place. There are existing techniques to do this (page faulting, various CPU tracing systems, access bit scanning etc) but they all have significant overheads.
- Monitoring accesses on the CXL device provides a path to getting good data without those overheads. These units are known as CXL Hotness Monitoring Units or CHMUs. Loosely speaking they count accesses to granuals of data (e.g. 4KiB pages). Exactly how they do that and where they sacrifice data accuracy is an implementation trade off.

Why do we need a model that gives real data?
- In general there is a need to develop software on top of these units to move data to the right place. Hard to evaluate that if we are making up the info on what is 'hot'.
- Need to allow for a bunch of 'impdef' solutions. Note that CHMU in this patch set is an oracle - it has enough counters to count every access. That's not realistic but it doesn't get me shouted at by our architecture teams for giving away any secrets. If we move forward with this, I'll probably implement a limited counter + full CAM solution (also unrealistic, but closer to real) I'd be very interested in contributions of other approaches (there are lots in the literature, under the term top-k)
- Resources will be constrained, so whilst a CHMU might in theory allow monitoring everything at once, that will come with a big accuracy cost. We need to design the algorithms that give us good data given those constraints.

So we need a solution to explore the design space and develop the software to take advantage of this hardware (there are various LSF/MM proposals on how to use this an other ways of tracking hotness).
https://lore.kernel.org/all/[email protected]/
https://lore.kernel.org/lkml/Z4XUoWlU-UgRik18@gourry-fedora-PF4VCD3F/T/

QEMU plugins give us a way to do this. In particular the existing Cache plugin can be easily modified to tell use what memory addresses missed at the last level of emulated cache. We can then filter those for the memory address range that maps to CXL and feed them to our counter implementation. On the other side, each instance of CXL type 3 device can connect to this server and request hotness monitoring services + provide parameters etc. Elements such as list threshold management and overflow detection etc are in the CXL HMU QEMU device mode. As noted above, we have an alternative approach that can closely couple things, so the device model registers the plugin directly and there is no server.

How to use it!
--------------

It runs a little slow but it runs and generates somewhat plausible outputs. I'd definitely suggest running it with the pass through optimization patch on the CXL staging tree (and a single direct connected device). Your millage will vary if you try to use other parameters, or hotness units beyond the first one (implementation far from complete!)

To run start the server in contrib/hmu/ providing a port number to listen on.

./chmu 4443

Then launch QEMU with something like the following.

qemu-system-aarch64 -icount shift=1 \
     -plugin ../qemu/bin/native/contrib/plugins/libcache.so,port=4443,missfilterbase=1099511627776,missfiltersize=1099511627776,dcachesize=8192,dassoc=4,dblksize=64,icachesize=8192,iassoc=4,iblksize=64,l2cachesize=32768,l2assoc=16,l2blksize=64 \
     -M virt,ras=on,nvdimm=on,gic-version=3,cxl=on,hmat=on -m 4g,maxmem=8g,slots=4 -cpu max -smp 4 \
     -kernel Image \
     -drive if=none,file=full.qcow2,format=qcow2,id=hd \
     -device pcie-root-port,id=root_port1 \
     -device virtio-blk-pci,drive=hd,x-max-bounce-buffer-size=512k \
     -nographic -no-reboot -append 'earlycon memblock=debug root=/dev/vda2 fsck.mode=skip maxcpus=4 tp_printk' \
     -monitor telnet:127.0.0.1:1234,server,nowait -bios QEMU_EFI.fd \
     -object memory-backend-ram,size=4G,id=mem0 \
     -object memory-backend-file,id=cxl-mem1,share=on,mem-path=/tmp/t3_cxl1.raw,size=1G,align=256M \
     -object memory-backend-file,id=cxl-mem2,share=on,mem-path=/tmp/t3_cxl2.raw,size=1G,align=256M \
     -object memory-backend-file,id=cxl-lsa1,share=on,mem-path=/tmp/t3_lsa1.raw,size=1M,align=1M \
      -object memory-backend-file,id=cxl-mem3,share=on,mem-path=/tmp/t3_cxl3.raw,size=1G,align=256M \
     -object memory-backend-file,id=cxl-mem4,share=on,mem-path=/tmp/t3_cxl4.raw,size=1G,align=256M \
     -object memory-backend-file,id=cxl-lsa2,share=on,mem-path=/tmp/t3_lsa2.raw,size=1M,align=1M \
     -device pxb-cxl,bus_nr=12,bus=pcie.0,id=cxl.1,hdm_for_passthrough=true,numa_node=0\
     -device cxl-rp,port=0,bus=cxl.1,id=cxl_rp_port0,chassis=0,slot=2 \
     -device cxl-type3,bus=cxl_rp_port0,volatile-memdev=cxl-mem1,id=cxl-pmem1,lsa=cxl-lsa1,sn=3,x-speed=32,x-width=16,chmu-port=4443 \
     -machine cxl-fmw.0.targets.0=cxl.1,cxl-fmw.0.size=8G,cxl-fmw.0.interleave-granularity=1k \
     -numa node,nodeid=0,cpus=0-3,memdev=mem0 \
     -numa node,nodeid=1 \
     -object acpi-generic-initiator,id=bob2,pci-dev=bob,node=1 \
     -numa node,nodeid=2 \
     -object acpi-generic-port,id=bob11,pci-bus=cxl.1,node=2 \

In the guest, create and bind the region - this brings up the CXL memory device so accesses go to the memory.

cd /sys/bus/cxl/devices/decoder0.0/
cat create_ram_region
echo region0 > create_ram_region
echo ram > /sys/bus/cxl/devices/decoder2.0/mode echo ram > /sys/bus/cxl/devices/decoder3.0/mode echo $((256 << 21)) > /sys/bus/cxl/devices/decoder2.0/dpa_size cd /sys/bus/cxl/devices/region0/
echo 256 > interleave_granularity
echo 1 > interleave_ways
echo $((256 << 21)) > size
echo decoder2.0 > target0
echo 1 > commit
echo region0 > /sys/bus/cxl/drivers/cxl_region/bind

Finally start perf with something like:

./perf record -a -e cxl_hmu_mem0.0.0/epoch_type=0,access_type=6,\
hotness_threshold=635,epoch_multiplier=4,epoch_scale=4,\
range_base=0,range_size=4096/ ./stress.sh

where stress.sh is

      sleep 2
      numactl --membind 3 stress-ng --vm 1 --vm-bytes 1M --vm-keep -t 5s
      sleep 2

See the results with
./perf report --dump-raw-trace | grep -A 200 HMU

Enjoy and have a good weekend.

Thanks,

Jonathan

Jonathan Cameron (3):
hw/cxl: Initial CXL Hotness Monitoring Unit Emulation
plugins: Add cache miss reporting over a socket.
contrib: Add example hotness monitoring unit server

include/hw/cxl/cxl.h | 1 +
include/hw/cxl/cxl_chmu.h | 154 ++++++++++++
include/hw/cxl/cxl_device.h | 13 +-
include/hw/cxl/cxl_pci.h | 7 +-
contrib/hmu/hmu.c | 312 ++++++++++++++++++++++++
contrib/plugins/cache.c | 75 +++++-
hw/cxl/cxl-chmu.c | 459 ++++++++++++++++++++++++++++++++++++
hw/mem/cxl_type3.c | 25 +-
hw/cxl/meson.build | 1 +
9 files changed, 1035 insertions(+), 12 deletions(-) create mode 100644 include/hw/cxl/cxl_chmu.h create mode 100644 contrib/hmu/hmu.c
create mode 100644 hw/cxl/cxl-chmu.c