Skip to contentSkip to conversation

xen/x86: Change stub page allocation/free

PATCH Superseded Roger Pau Monne 3 participants 4 messages Jun 11

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 xen/x86: Change stub page allocation/free

From: Jason Andryuk <[email protected]>

Today the inline tracking of the stub page is problematic. 0xcc is used to indicate unused, but it is also a "clear value." A !CONFIG_PV build with smt=0 will bring up CPU0, bring up CPU1, bring down CPU1, and free the in-use stub page. CPU0 or subsequent onlined CPUs can write to the re-used page.

The new approach uses a global, CPU-indexed dynamically allocated array of stub addresses. However, to handle NUMA aware allocations, we cannot allocate all the memory in advance because of the NUMA dependency. Take advantage of the fact that Xen will attempt to contiguously pack CPUs on the same NUMA node (see normalise_cpu_order()), and on CPU bringup use the same stubs page the previous CPU did if suitable. Note the code would still function properly even if CPUs from NUMA nodes are not contiguously packed, it just consumes more memory.

stub pages are no longer freed. They remain referenced in the global CPU-indexed array and are re-used if the CPU is re-onlined.

stubs and node_stubs don't have an explicit lock. During boot they are accessed single threaded. During runtime, &cpu_add_remove_lock serializes access.

Fixes 7a66ac8d1633 ("x86: move syscall trampolines off the stack")

---
There are other even more simple options here: for example Andrew proposed to pack stubs contiguously in both the physical and the linear address spaces, at the cost of possibly loosing the NUMA memory affinity between the allocated page and the CPU using it. We have decided to go for a more conservative approach here, that keeps the same properties as the current logic regarding NUMA memory affinity of the stub region.

---
 xen/arch/x86/include/asm/stubs.h |  2 +-
 xen/arch/x86/setup.c             |  4 +-
 xen/arch/x86/smpboot.c           | 91 +++++++++++++++++---------------
 3 files changed, 49 insertions(+), 48 deletions(-)
xen/arch/x86/include/asm/stubs.h +1 -1
index a520928e9a50..7d8d302e0623 100644
@@ -32,6 +32,6 @@ struct stubs {
 };
 
 DECLARE_PER_CPU(struct stubs, stubs);
-unsigned long alloc_stub_page(unsigned int cpu, unsigned long *mfn);
+void init_bsp_stub(void);
 
 #endif /* X86_ASM_STUBS_H */
xen/arch/x86/setup.c +1 -3
index 4192edf635b6..cddf8806c877 100644
@@ -2089,9 +2089,7 @@ void asmlinkage __init noreturn __start_xen(void)
 
     init_idle_domain();
 
-    this_cpu(stubs.addr) = alloc_stub_page(smp_processor_id(),
-                                           &this_cpu(stubs).mfn);
-    BUG_ON(!this_cpu(stubs.addr));
+    init_bsp_stub();
 
     bsp_traps_reinit(); /* Needs stubs allocated, must be before presmp_initcalls. */
 
xen/arch/x86/smpboot.c +47 -44

141 lines not on this page · raw message

Signature · 1 line
2.53.0
#1 lore raw
Quoted 7 lines · Roger Pau Monne

From: Jason Andryuk <[email protected]>

Today the inline tracking of the stub page is problematic. 0xcc is used to indicate unused, but it is also a "clear value." A !CONFIG_PV build with smt=0 will bring up CPU0, bring up CPU1, bring down CPU1, and free the in-use stub page. CPU0 or subsequent onlined CPUs can write to the re-used page.

I'm pretty sure a CONFIG_PV build booted on a FRED enabled system will do the same.

This is the other case where we (now) forgo writing out the LSTAR/CSTAR stubs.

Quoted 10 lines

The new approach uses a global, CPU-indexed dynamically allocated array of stub addresses. However, to handle NUMA aware allocations, we cannot allocate all the memory in advance because of the NUMA dependency. Take advantage of the fact that Xen will attempt to contiguously pack CPUs on the same NUMA node (see normalise_cpu_order()), and on CPU bringup use the same stubs page the previous CPU did if suitable. Note the code would still function properly even if CPUs from NUMA nodes are not contiguously packed, it just consumes more memory.

Quoted 6 lines

stub pages are no longer freed. They remain referenced in the global CPU-indexed array and are re-used if the CPU is re-onlined.

stubs and node_stubs don't have an explicit lock. During boot they are accessed single threaded. During runtime, &cpu_add_remove_lock serializes access.

Is node_stubs stale?  Stub(s) should be capitalised at the start of the sentence.  In context, it's not clear that it's a variable name, and it doesn't need to be the literal variable name to convey the intended meaning.

Quoted 11 lines
Fixes 7a66ac8d1633 ("x86: move syscall trampolines off the stack")

---
There are other even more simple options here: for example Andrew proposed to pack stubs contiguously in both the physical and the linear address spaces, at the cost of possibly loosing the NUMA memory affinity between the allocated page and the CPU using it. We have decided to go for a more conservative approach here, that keeps the same properties as the current logic regarding NUMA memory affinity of the stub region.

Part of the suggestion was made in error, but there's a second aspect which I'll discuss at the end of the email.

xen/arch/x86/setup.c void asmlinkage __init noreturn __start_xen(void) · line 2092

Quoted 12 lines
xen/arch/x86/setup.c +1 -3
index 4192edf635b6..cddf8806c877 100644
@@ -2089,9 +2089,7 @@ void asmlinkage __init noreturn __start_xen(void)
 
     init_idle_domain();
 
-    this_cpu(stubs.addr) = alloc_stub_page(smp_processor_id(),
-                                           &this_cpu(stubs).mfn);
-    BUG_ON(!this_cpu(stubs.addr));
+    init_bsp_stub();

Personally, I'd name this init_stubs().  It does work for more than just the BSP, and the bsp_* part is only really needed for clarity when there's a matching ap_* variant, which is not the case here.

Quoted 21 lines
xen/arch/x86/setup.c +0 -0
     bsp_traps_reinit(); /* Needs stubs allocated, must be before presmp_initcalls. */
 
xen/arch/x86/smpboot.c +4 -3
index d8fd71ffab37..3282392317f4 100644
@@ -641,41 +642,61 @@ static int do_boot_cpu(int apicid, int cpu)
     return rc;
 }
 
-#define STUB_BUF_CPU_OFFS(cpu) (((cpu) & (STUBS_PER_PAGE - 1)) * STUB_BUF_SIZE)
+/* Dynamically allocated, indexed by CPU.  Store physical address of stubs. */
+static paddr_t *__ro_after_init stubs;
 
-unsigned long alloc_stub_page(unsigned int cpu, unsigned long *mfn)
+static bool assign_stub_page(unsigned int cpu)
 {
     unsigned long stub_va;
-    struct page_info *pg;
+    paddr_t addr = stubs[cpu];

xen/arch/x86/smpboot.c static int do_boot_cpu(int apicid, int cpu) · line 658

Quoted 10 lines
xen/arch/x86/smpboot.c +1 -7
 
-    BUILD_BUG_ON(STUBS_PER_PAGE & (STUBS_PER_PAGE - 1));
-
-    if ( *mfn )
-        pg = mfn_to_page(_mfn(*mfn));
-    else
+    if ( addr == INVALID_PADDR )
     {
-        nodeid_t node = cpu_to_node(cpu);
-        unsigned int memflags = node != NUMA_NO_NODE ? MEMF_node(node) : 0;

I think you need to retain this calculation of memflags. MEMF_node(NUMA_NO_NODE) doesn't work as expected.

Quoted 7 lines
xen/arch/x86/smpboot.c +3 -3
+        nodeid_t nid = cpu_to_node(cpu);
 
-        pg = alloc_domheap_page(NULL, memflags);
-        if ( !pg )
-            return 0;
+        /*
+         * Attempt to use the same page as the previous CPU if possible,

xen/arch/x86/smpboot.c static int do_boot_cpu(int apicid, int cpu) · line 664

Quoted 10 lines
xen/arch/x86/smpboot.c +8 -1
+         * otherwise allocate a new one.
+         */
+        if ( cpu && nid == cpu_to_node(cpu - 1) &&
+             PAGE_OFFSET(stubs[cpu - 1] + STUB_BUF_SIZE) )
+            addr = stubs[cpu - 1] + STUB_BUF_SIZE;
+        else
+        {
+            struct page_info *pg = alloc_domheap_page(NULL, MEMF_node(nid));
 
-        unmap_domain_page(memset(__map_domain_page(pg), 0xcc, PAGE_SIZE));

You've dropped this memset() of the whole page to 0xcc.

As a consequence, the stubs for not-yet-onlined CPUs, or for gaps because of NUMA, are rubble yet mapped executably.

Quoted 31 lines
xen/arch/x86/smpboot.c +17 -9
+            if ( !pg )
+                return false;
+            addr = page_to_maddr(pg);
+        }
+        stubs[cpu] = addr;
     }
 
     stub_va = XEN_VIRT_END - FIXADDR_X_SIZE - (cpu + 1) * PAGE_SIZE;
-    if ( map_pages_to_xen(stub_va, page_to_mfn(pg), 1,
+    if ( map_pages_to_xen(stub_va, maddr_to_mfn(addr), 1,
                           PAGE_HYPERVISOR_RX | MAP_SMALL_PAGES) )
-    {
-        if ( !*mfn )
-            free_domheap_page(pg);
-        stub_va = 0;
-    }
-    else if ( !*mfn )
-        *mfn = mfn_x(page_to_mfn(pg));
+        return false;
 
-    return stub_va;
+    per_cpu(stubs.mfn, cpu) = PFN_DOWN(addr);
+    per_cpu(stubs.addr, cpu) = stub_va + PAGE_OFFSET(addr);
+    return true;
+}
+
+void __init init_bsp_stub(void)
+{
+    const unsigned int num_cpus = num_present_cpus();
+    unsigned int i;
+

xen/arch/x86/smpboot.c static int do_boot_cpu(int apicid, int cpu) · line 699

Quoted 10 lines
xen/arch/x86/smpboot.c +10 -0
+    ASSERT(!stubs);
+    stubs = xvmalloc_array(typeof(*stubs), num_cpus);
+    if ( !stubs )
+        panic("Unable to allocate stub array");
+
+    for ( i = 0; i < num_cpus; i++ )
+        stubs[i] = INVALID_PADDR;
+
+    if ( !assign_stub_page(0) )
+        panic("Unable to initialize BSP stub region");

\n's for both panic messages.

With the above stuff addressed, I think this is looking ok, but definitely subject to Jason confirming it resolves his issue.  And for 4.22, that might even be sufficient to go in.

The other thing I want to discuss is this:

xen/arch/x86/smpboot.c static int do_boot_cpu(int apicid, int cpu) · line 675

Quoted 1 line
xen/arch/x86/smpboot.c +0 -0
     stub_va = XEN_VIRT_END - FIXADDR_X_SIZE - (cpu + 1) * PAGE_SIZE;

because it creates 32 virtual aliases of every stub.

AIUI, this was a hard requirement for the old freeing scheme, but the optimisation guides recommend against creating aliases like this. Besides microarchitectural tracking/safety effects, one consequence is that we end up with 31 aliases which have unsafe branches in them; disp32's depend on the linear address the code is executed at.

The ideal solution would be allocate VAs just like we allocate paddrs, and for the map_pages_to_xen() be beside the alloc_domheap_page(), rather than outside of INVALID_PADDR check.

This reduces the amount of VA space (and L1 pagetables) used by 32 times, and removes the risk of accidentally using the wrong alias.

~Andrew

#2 lore raw
Quoted 170 lines · Andrew Cooper
Quoted 7 lines · Roger Pau Monne

From: Jason Andryuk <[email protected]>

Today the inline tracking of the stub page is problematic. 0xcc is used to indicate unused, but it is also a "clear value." A !CONFIG_PV build with smt=0 will bring up CPU0, bring up CPU1, bring down CPU1, and free the in-use stub page. CPU0 or subsequent onlined CPUs can write to the re-used page.

I'm pretty sure a CONFIG_PV build booted on a FRED enabled system will do the same.

This is the other case where we (now) forgo writing out the LSTAR/CSTAR stubs.

Quoted 10 lines

The new approach uses a global, CPU-indexed dynamically allocated array of stub addresses. However, to handle NUMA aware allocations, we cannot allocate all the memory in advance because of the NUMA dependency. Take advantage of the fact that Xen will attempt to contiguously pack CPUs on the same NUMA node (see normalise_cpu_order()), and on CPU bringup use the same stubs page the previous CPU did if suitable. Note the code would still function properly even if CPUs from NUMA nodes are not contiguously packed, it just consumes more memory.

Quoted 6 lines

stub pages are no longer freed. They remain referenced in the global CPU-indexed array and are re-used if the CPU is re-onlined.

stubs and node_stubs don't have an explicit lock. During boot they are accessed single threaded. During runtime, &cpu_add_remove_lock serializes access.

Is node_stubs stale?  Stub(s) should be capitalised at the start of the sentence.  In context, it's not clear that it's a variable name, and it doesn't need to be the literal variable name to convey the intended meaning.

Quoted 11 lines
Fixes 7a66ac8d1633 ("x86: move syscall trampolines off the stack")

---
There are other even more simple options here: for example Andrew proposed to pack stubs contiguously in both the physical and the linear address spaces, at the cost of possibly loosing the NUMA memory affinity between the allocated page and the CPU using it. We have decided to go for a more conservative approach here, that keeps the same properties as the current logic regarding NUMA memory affinity of the stub region.

Part of the suggestion was made in error, but there's a second aspect which I'll discuss at the end of the email.

xen/arch/x86/setup.c void asmlinkage __init noreturn __start_xen(void) · line 2092

Quoted 12 lines
xen/arch/x86/setup.c +1 -3
index 4192edf635b6..cddf8806c877 100644
@@ -2089,9 +2089,7 @@ void asmlinkage __init noreturn __start_xen(void)
 
      init_idle_domain();
 
-    this_cpu(stubs.addr) = alloc_stub_page(smp_processor_id(),
-                                           &this_cpu(stubs).mfn);
-    BUG_ON(!this_cpu(stubs.addr));
+    init_bsp_stub();

Personally, I'd name this init_stubs().  It does work for more than just the BSP, and the bsp_* part is only really needed for clarity when there's a matching ap_* variant, which is not the case here.

Quoted 21 lines
xen/arch/x86/setup.c +0 -0
      bsp_traps_reinit(); /* Needs stubs allocated, must be before presmp_initcalls. */
 
xen/arch/x86/smpboot.c +4 -3
index d8fd71ffab37..3282392317f4 100644
@@ -641,41 +642,61 @@ static int do_boot_cpu(int apicid, int cpu)
      return rc;
  }
 
-#define STUB_BUF_CPU_OFFS(cpu) (((cpu) & (STUBS_PER_PAGE - 1)) * STUB_BUF_SIZE)
+/* Dynamically allocated, indexed by CPU.  Store physical address of stubs. */
+static paddr_t *__ro_after_init stubs;
 
-unsigned long alloc_stub_page(unsigned int cpu, unsigned long *mfn)
+static bool assign_stub_page(unsigned int cpu)
  {
      unsigned long stub_va;
-    struct page_info *pg;
+    paddr_t addr = stubs[cpu];

xen/arch/x86/smpboot.c static int do_boot_cpu(int apicid, int cpu) · line 658

Quoted 10 lines
xen/arch/x86/smpboot.c +1 -7
 
-    BUILD_BUG_ON(STUBS_PER_PAGE & (STUBS_PER_PAGE - 1));
-
-    if ( *mfn )
-        pg = mfn_to_page(_mfn(*mfn));
-    else
+    if ( addr == INVALID_PADDR )
      {
-        nodeid_t node = cpu_to_node(cpu);
-        unsigned int memflags = node != NUMA_NO_NODE ? MEMF_node(node) : 0;

I think you need to retain this calculation of memflags. MEMF_node(NUMA_NO_NODE) doesn't work as expected.

Quoted 7 lines
xen/arch/x86/smpboot.c +3 -3
+        nodeid_t nid = cpu_to_node(cpu);
 
-        pg = alloc_domheap_page(NULL, memflags);
-        if ( !pg )
-            return 0;
+        /*
+         * Attempt to use the same page as the previous CPU if possible,

xen/arch/x86/smpboot.c static int do_boot_cpu(int apicid, int cpu) · line 664

Quoted 10 lines
xen/arch/x86/smpboot.c +8 -1
+         * otherwise allocate a new one.
+         */
+        if ( cpu && nid == cpu_to_node(cpu - 1) &&
+             PAGE_OFFSET(stubs[cpu - 1] + STUB_BUF_SIZE) )
+            addr = stubs[cpu - 1] + STUB_BUF_SIZE;
+        else
+        {
+            struct page_info *pg = alloc_domheap_page(NULL, MEMF_node(nid));
 
-        unmap_domain_page(memset(__map_domain_page(pg), 0xcc, PAGE_SIZE));

You've dropped this memset() of the whole page to 0xcc.

As a consequence, the stubs for not-yet-onlined CPUs, or for gaps because of NUMA, are rubble yet mapped executably.

Quoted 31 lines
xen/arch/x86/smpboot.c +17 -9
+            if ( !pg )
+                return false;
+            addr = page_to_maddr(pg);
+        }
+        stubs[cpu] = addr;
      }
 
      stub_va = XEN_VIRT_END - FIXADDR_X_SIZE - (cpu + 1) * PAGE_SIZE;
-    if ( map_pages_to_xen(stub_va, page_to_mfn(pg), 1,
+    if ( map_pages_to_xen(stub_va, maddr_to_mfn(addr), 1,
                            PAGE_HYPERVISOR_RX | MAP_SMALL_PAGES) )
-    {
-        if ( !*mfn )
-            free_domheap_page(pg);
-        stub_va = 0;
-    }
-    else if ( !*mfn )
-        *mfn = mfn_x(page_to_mfn(pg));
+        return false;
 
-    return stub_va;
+    per_cpu(stubs.mfn, cpu) = PFN_DOWN(addr);
+    per_cpu(stubs.addr, cpu) = stub_va + PAGE_OFFSET(addr);
+    return true;
+}
+
+void __init init_bsp_stub(void)
+{
+    const unsigned int num_cpus = num_present_cpus();
+    unsigned int i;
+

xen/arch/x86/smpboot.c static int do_boot_cpu(int apicid, int cpu) · line 699

Quoted 10 lines
xen/arch/x86/smpboot.c +10 -0
+    ASSERT(!stubs);
+    stubs = xvmalloc_array(typeof(*stubs), num_cpus);
+    if ( !stubs )
+        panic("Unable to allocate stub array");
+
+    for ( i = 0; i < num_cpus; i++ )
+        stubs[i] = INVALID_PADDR;
+
+    if ( !assign_stub_page(0) )
+        panic("Unable to initialize BSP stub region");

\n's for both panic messages.

Quoted 3 lines

With the above stuff addressed, I think this is looking ok, but definitely subject to Jason confirming it resolves his issue.  And for 4.22, that might even be sufficient to go in.

Thanks.

~ Oleksii

Quoted 21 lines

The other thing I want to discuss is this:

xen/arch/x86/smpboot.c static int do_boot_cpu(int apicid, int cpu) · line 675

Quoted 1 line
xen/arch/x86/smpboot.c +0 -0
      stub_va = XEN_VIRT_END - FIXADDR_X_SIZE - (cpu + 1) * PAGE_SIZE;

because it creates 32 virtual aliases of every stub.

AIUI, this was a hard requirement for the old freeing scheme, but the optimisation guides recommend against creating aliases like this. Besides microarchitectural tracking/safety effects, one consequence is that we end up with 31 aliases which have unsafe branches in them; disp32's depend on the linear address the code is executed at.

The ideal solution would be allocate VAs just like we allocate paddrs, and for the map_pages_to_xen() be beside the alloc_domheap_page(), rather than outside of INVALID_PADDR check.

This reduces the amount of VA space (and L1 pagetables) used by 32 times, and removes the risk of accidentally using the wrong alias.

~Andrew

#3 lore raw
In reply to Andrew Cooper
Quoted 28 lines · Andrew Cooper
Quoted 7 lines · Roger Pau Monne

From: Jason Andryuk <[email protected]>

Today the inline tracking of the stub page is problematic. 0xcc is used to indicate unused, but it is also a "clear value." A !CONFIG_PV build with smt=0 will bring up CPU0, bring up CPU1, bring down CPU1, and free the in-use stub page. CPU0 or subsequent onlined CPUs can write to the re-used page.

I'm pretty sure a CONFIG_PV build booted on a FRED enabled system will do the same.

This is the other case where we (now) forgo writing out the LSTAR/CSTAR stubs.

Quoted 13 lines

The new approach uses a global, CPU-indexed dynamically allocated array of stub addresses. However, to handle NUMA aware allocations, we cannot allocate all the memory in advance because of the NUMA dependency. Take advantage of the fact that Xen will attempt to contiguously pack CPUs on the same NUMA node (see normalise_cpu_order()), and on CPU bringup use the same stubs page the previous CPU did if suitable. Note the code would still function properly even if CPUs from NUMA nodes are not contiguously packed, it just consumes more memory.

stub pages are no longer freed. They remain referenced in the global CPU-indexed array and are re-used if the CPU is re-onlined.

Quoted 7 lines
Quoted 3 lines

stubs and node_stubs don't have an explicit lock. During boot they are accessed single threaded. During runtime, &cpu_add_remove_lock serializes access.

Is node_stubs stale?  Stub(s) should be capitalised at the start of the sentence.  In context, it's not clear that it's a variable name, and it doesn't need to be the literal variable name to convey the intended meaning.

Oh, yes, that's a leftover from the previous version that I picked up from Jason. What about using "The stubs array don't have..." to make it clear it's referring to the variable.

Quoted 23 lines
Quoted 11 lines
Fixes 7a66ac8d1633 ("x86: move syscall trampolines off the stack")

---
There are other even more simple options here: for example Andrew proposed to pack stubs contiguously in both the physical and the linear address spaces, at the cost of possibly loosing the NUMA memory affinity between the allocated page and the CPU using it. We have decided to go for a more conservative approach here, that keeps the same properties as the current logic regarding NUMA memory affinity of the stub region.

Part of the suggestion was made in error, but there's a second aspect which I'll discuss at the end of the email.

Quoted 8 lines
xen/arch/x86/setup.c +0 -0
index 4192edf635b6..cddf8806c877 100644
@@ -2089,9 +2089,7 @@ void asmlinkage __init noreturn __start_xen(void)
 
     init_idle_domain();
 
Quoted 8 lines

xen/arch/x86/setup.c void asmlinkage __init noreturn __start_xen(void) · line 2092

Quoted 4 lines
xen/arch/x86/setup.c +1 -3
-    this_cpu(stubs.addr) = alloc_stub_page(smp_processor_id(),
-                                           &this_cpu(stubs).mfn);
-    BUG_ON(!this_cpu(stubs.addr));
+    init_bsp_stub();

Personally, I'd name this init_stubs().  It does work for more than just the BSP, and the bsp_* part is only really needed for clarity when there's a matching ap_* variant, which is not the case here.

Np, I will change it.

Quoted 22 lines
Quoted 22 lines
xen/arch/x86/setup.c +0 -0
     bsp_traps_reinit(); /* Needs stubs allocated, must be before presmp_initcalls. */
 
xen/arch/x86/smpboot.c +4 -3
index d8fd71ffab37..3282392317f4 100644
@@ -641,41 +642,61 @@ static int do_boot_cpu(int apicid, int cpu)
     return rc;
 }
 
-#define STUB_BUF_CPU_OFFS(cpu) (((cpu) & (STUBS_PER_PAGE - 1)) * STUB_BUF_SIZE)
+/* Dynamically allocated, indexed by CPU.  Store physical address of stubs. */
+static paddr_t *__ro_after_init stubs;
 
-unsigned long alloc_stub_page(unsigned int cpu, unsigned long *mfn)
+static bool assign_stub_page(unsigned int cpu)
 {
     unsigned long stub_va;
-    struct page_info *pg;
+    paddr_t addr = stubs[cpu];
 
Quoted 12 lines

xen/arch/x86/smpboot.c static int do_boot_cpu(int apicid, int cpu) · line 658

Quoted 9 lines
xen/arch/x86/smpboot.c +1 -7
-    BUILD_BUG_ON(STUBS_PER_PAGE & (STUBS_PER_PAGE - 1));
-
-    if ( *mfn )
-        pg = mfn_to_page(_mfn(*mfn));
-    else
+    if ( addr == INVALID_PADDR )
     {
-        nodeid_t node = cpu_to_node(cpu);
-        unsigned int memflags = node != NUMA_NO_NODE ? MEMF_node(node) : 0;

I think you need to retain this calculation of memflags. MEMF_node(NUMA_NO_NODE) doesn't work as expected.

I think I'm confused, MEMF_node() macro is:

((((n) + 1) & MEMF_node_mask) << _MEMF_node)

So when n == 0xff the memflags result is 0 (0x100 & 0xff). And then doing MEMF_get_node(0) gives you NUMA_NO_NODE, so works as expected?

Quoted 16 lines
Quoted 16 lines
xen/arch/x86/smpboot.c +11 -3
+        nodeid_t nid = cpu_to_node(cpu);
 
-        pg = alloc_domheap_page(NULL, memflags);
-        if ( !pg )
-            return 0;
+        /*
+         * Attempt to use the same page as the previous CPU if possible,
+         * otherwise allocate a new one.
+         */
+        if ( cpu && nid == cpu_to_node(cpu - 1) &&
+             PAGE_OFFSET(stubs[cpu - 1] + STUB_BUF_SIZE) )
+            addr = stubs[cpu - 1] + STUB_BUF_SIZE;
+        else
+        {
+            struct page_info *pg = alloc_domheap_page(NULL, MEMF_node(nid));
 
Quoted 6 lines

xen/arch/x86/smpboot.c static int do_boot_cpu(int apicid, int cpu) · line 664

Quoted 1 line
xen/arch/x86/smpboot.c +0 -1
-        unmap_domain_page(memset(__map_domain_page(pg), 0xcc, PAGE_SIZE));

You've dropped this memset() of the whole page to 0xcc.

As a consequence, the stubs for not-yet-onlined CPUs, or for gaps because of NUMA, are rubble yet mapped executably.

My bad, sorry, I will re-add it.

Quoted 44 lines
Quoted 31 lines
xen/arch/x86/smpboot.c +17 -9
+            if ( !pg )
+                return false;
+            addr = page_to_maddr(pg);
+        }
+        stubs[cpu] = addr;
     }
 
     stub_va = XEN_VIRT_END - FIXADDR_X_SIZE - (cpu + 1) * PAGE_SIZE;
-    if ( map_pages_to_xen(stub_va, page_to_mfn(pg), 1,
+    if ( map_pages_to_xen(stub_va, maddr_to_mfn(addr), 1,
                           PAGE_HYPERVISOR_RX | MAP_SMALL_PAGES) )
-    {
-        if ( !*mfn )
-            free_domheap_page(pg);
-        stub_va = 0;
-    }
-    else if ( !*mfn )
-        *mfn = mfn_x(page_to_mfn(pg));
+        return false;
 
-    return stub_va;
+    per_cpu(stubs.mfn, cpu) = PFN_DOWN(addr);
+    per_cpu(stubs.addr, cpu) = stub_va + PAGE_OFFSET(addr);
+    return true;
+}
+
+void __init init_bsp_stub(void)
+{
+    const unsigned int num_cpus = num_present_cpus();
+    unsigned int i;
+

xen/arch/x86/smpboot.c static int do_boot_cpu(int apicid, int cpu) · line 699

Quoted 10 lines
xen/arch/x86/smpboot.c +10 -0
+    ASSERT(!stubs);
+    stubs = xvmalloc_array(typeof(*stubs), num_cpus);
+    if ( !stubs )
+        panic("Unable to allocate stub array");
+
+    for ( i = 0; i < num_cpus; i++ )
+        stubs[i] = INVALID_PADDR;
+
+    if ( !assign_stub_page(0) )
+        panic("Unable to initialize BSP stub region");

\n's for both panic messages.

Quoted 3 lines

With the above stuff addressed, I think this is looking ok, but definitely subject to Jason confirming it resolves his issue.  And for 4.22, that might even be sufficient to go in.

I will go with the adjusted version of this patch for 4.22, if it fixes Jason's issue.

Quoted 12 lines

The other thing I want to discuss is this:

xen/arch/x86/smpboot.c static int do_boot_cpu(int apicid, int cpu) · line 675

Quoted 1 line
xen/arch/x86/smpboot.c +0 -0
     stub_va = XEN_VIRT_END - FIXADDR_X_SIZE - (cpu + 1) * PAGE_SIZE;

because it creates 32 virtual aliases of every stub.

AIUI, this was a hard requirement for the old freeing scheme, but the optimisation guides recommend against creating aliases like this. Besides microarchitectural tracking/safety effects, one consequence is that we end up with 31 aliases which have unsafe branches in them; disp32's depend on the linear address the code is executed at.

Quoted 6 lines

The ideal solution would be allocate VAs just like we allocate paddrs, and for the map_pages_to_xen() be beside the alloc_domheap_page(), rather than outside of INVALID_PADDR check.

This reduces the amount of VA space (and L1 pagetables) used by 32 times, and removes the risk of accidentally using the wrong alias.

I think the point of having a different page for each CPU is that we can remove the page-table entry when the CPU goes offline, and hence any mismatched attempts to use that stub region by a different CPU would result in a page-fault. In practice I'm not sure how useful that is, at the end we already clobber the stubs region with int3's.

I don't mind looking into that as a followup change, but for 4.22 I would rather take this less risky version. The mapping adjustments will be a different patch anyway, I don't think it should be squashed with the content here.

Thanks, Roger.