[PATCHv5 1/2] powerpc/pseries: group lmb operation and memblock's

This patch prepares for the incoming patch which swaps the order of
KOBJ_ADD/REMOVE uevent and dt's updating.

The dt updating should come after lmb operations, and before
__remove_memory()/__add_memory().  Accordingly, grouping all lmb operations
before the memblock's.

Signed-off-by: Pingfan Liu <redacted>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Hari Bathini <hbathini@linux.ibm.com>
Cc: Nathan Lynch <redacted>
Cc: Nathan Fontenot <redacted>
Cc: Laurent Dufour <redacted>
To: linuxppc-dev@lists.ozlabs.org
Cc: kexec@lists.infradead.org
---
v4 -> v5: fix the miss of clearing DRCONF_MEM_ASSIGNED in a failure path
 arch/powerpc/platforms/pseries/hotplug-memory.c | 28 +++++++++++++++++--------
 1 file changed, 19 insertions(+), 9 deletions(-)

diff --git a/arch/powerpc/platforms/pseries/hotplug-memory.c b/arch/powerpc/platforms/pseries/hotplug-memory.c
index 5d545b7..46cbcd1 100644
--- a/arch/powerpc/platforms/pseries/hotplug-memory.c
+++ b/arch/powerpc/platforms/pseries/hotplug-memory.c

@@ -355,7 +355,8 @@ static int dlpar_add_lmb(struct drmem_lmb *);
 static int dlpar_remove_lmb(struct drmem_lmb *lmb)
 {
 	unsigned long block_sz;
-	int rc;
+	phys_addr_t base_addr;
+	int rc, nid;
 
 	if (!lmb_is_removable(lmb))
 		return -EINVAL;

@@ -364,17 +365,19 @@ static int dlpar_remove_lmb(struct drmem_lmb *lmb)
 	if (rc)
 		return rc;
 
+	base_addr = lmb->base_addr;
+	nid = lmb->nid;
 	block_sz = pseries_memory_block_size();
 
-	__remove_memory(lmb->nid, lmb->base_addr, block_sz);
-
-	/* Update memory regions for memory remove */
-	memblock_remove(lmb->base_addr, block_sz);
-
 	invalidate_lmb_associativity_index(lmb);
 	lmb_clear_nid(lmb);
 	lmb->flags &= ~DRCONF_MEM_ASSIGNED;
 
+	__remove_memory(nid, base_addr, block_sz);
+
+	/* Update memory regions for memory remove */
+	memblock_remove(base_addr, block_sz);
+
 	return 0;
 }
 

@@ -603,22 +606,29 @@ static int dlpar_add_lmb(struct drmem_lmb *lmb)
 	}
 
 	lmb_set_nid(lmb);
+	lmb->flags |= DRCONF_MEM_ASSIGNED;
+
 	block_sz = memory_block_size_bytes();
 
 	/* Add the memory */
 	rc = __add_memory(lmb->nid, lmb->base_addr, block_sz);
 	if (rc) {
 		invalidate_lmb_associativity_index(lmb);
+		lmb_clear_nid(lmb);
+		lmb->flags &= ~DRCONF_MEM_ASSIGNED;
 		return rc;
 	}
 
 	rc = dlpar_online_lmb(lmb);
 	if (rc) {
-		__remove_memory(lmb->nid, lmb->base_addr, block_sz);
+		int nid = lmb->nid;
+		phys_addr_t base_addr = lmb->base_addr;
+
 		invalidate_lmb_associativity_index(lmb);
 		lmb_clear_nid(lmb);
-	} else {
-		lmb->flags |= DRCONF_MEM_ASSIGNED;
+		lmb->flags &= ~DRCONF_MEM_ASSIGNED;
+
+		__remove_memory(nid, base_addr, block_sz);
 	}
 
 	return rc;

-- 
2.7.5

A bug is observed on pseries by taking the following steps on rhel:
-1. drmgr -c mem -r -q 5
-2. echo c > /proc/sysrq-trigger

And then, the failure looks like:
kdump: saving to /sysroot//var/crash/127.0.0.1-2020-01-16-02:06:14/
kdump: saving vmcore-dmesg.txt
kdump: saving vmcore-dmesg.txt complete
kdump: saving vmcore
 Checking for memory holes                         : [  0.0 %] /                   Checking for memory holes                         : [100.0 %] |                   Excluding unnecessary pages                       : [100.0 %] \                   Copying data                                      : [  0.3 %] -          eta: 38s[   44.337636] hash-mmu: mm: Hashing failure ! EA=0x7fffba400000 access=0x8000000000000004 current=makedumpfile
[   44.337663] hash-mmu:     trap=0x300 vsid=0x13a109c ssize=1 base psize=2 psize 2 pte=0xc000000050000504
[   44.337677] hash-mmu: mm: Hashing failure ! EA=0x7fffba400000 access=0x8000000000000004 current=makedumpfile
[   44.337692] hash-mmu:     trap=0x300 vsid=0x13a109c ssize=1 base psize=2 psize 2 pte=0xc000000050000504
[   44.337708] makedumpfile[469]: unhandled signal 7 at 00007fffba400000 nip 00007fffbbc4d7fc lr 000000011356ca3c code 2
[   44.338548] Core dump to |/bin/false pipe failed
/lib/kdump-lib-initramfs.sh: line 98:   469 Bus error               $CORE_COLLECTOR /proc/vmcore $_mp/$KDUMP_PATH/$HOST_IP-$DATEDIR/vmcore-incomplete
kdump: saving vmcore failed

* Root cause *
  After analyzing, it turns out that in the current implementation,
when hot-removing lmb, the KOBJ_REMOVE event ejects before the dt updating as
the code __remove_memory() comes before drmem_update_dt().
So in kdump kernel, when read_from_oldmem() resorts to
pSeries_lpar_hpte_insert() to install hpte, but fails with -2 due to
non-exist pfn. And finally, low_hash_fault() raise SIGBUS to process, as it
can be observed "Bus error"

From a viewpoint of listener and publisher, the publisher notifies the
listener before data is ready.  This introduces a problem where udev
launches kexec-tools (due to KOBJ_REMOVE) and loads a stale dt before
updating. And in capture kernel, makedumpfile will access the memory based
on the stale dt info, and hit a SIGBUS error due to an un-existed lmb.

* Fix *
This bug is introduced by commit 063b8b1251fd
("powerpc/pseries/memory-hotplug: Only update DT once per memory DLPAR
request"), which tried to combine all the dt updating into one.

To fix this issue, meanwhile not to introduce a quadratic runtime
complexity by the model:
  dlpar_memory_add_by_count
    for_each_drmem_lmb             <--
      dlpar_add_lmb
        drmem_update_dt(_v1|_v2)
          for_each_drmem_lmb       <--
The dt should still be only updated once, and just before the last memory
online/offline event is ejected to user space. Achieve this by tracing the
num of lmb added or removed.

Signed-off-by: Pingfan Liu <redacted>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Hari Bathini <hbathini@linux.ibm.com>
Cc: Nathan Lynch <redacted>
Cc: Nathan Fontenot <redacted>
Cc: Laurent Dufour <redacted>
To: linuxppc-dev@lists.ozlabs.org
Cc: kexec@lists.infradead.org
---
v4 -> v5: change dlpar_add_lmb()/dlpar_remove_lmb() prototype to report
          whether dt is updated successfully.
          Fix a condition boundary check bug
v3 -> v4: resolve a quadratic runtime complexity issue.
          This series is applied on next-test branch
 arch/powerpc/platforms/pseries/hotplug-memory.c | 102 +++++++++++++++++++-----
 1 file changed, 80 insertions(+), 22 deletions(-)

diff --git a/arch/powerpc/platforms/pseries/hotplug-memory.c b/arch/powerpc/platforms/pseries/hotplug-memory.c
index 46cbcd1..1567d9f 100644
--- a/arch/powerpc/platforms/pseries/hotplug-memory.c
+++ b/arch/powerpc/platforms/pseries/hotplug-memory.c

@@ -350,13 +350,22 @@ static bool lmb_is_removable(struct drmem_lmb *lmb)
 	return true;
 }
 
-static int dlpar_add_lmb(struct drmem_lmb *);
+enum dt_update_status {
+	DT_NOUPDATE,
+	DT_TOUPDATE,
+	DT_UPDATED,
+};
+
+/* "*dt_update" returns DT_UPDATED if updated */
+static int dlpar_add_lmb(struct drmem_lmb *lmb,
+		enum dt_update_status *dt_update);
 
-static int dlpar_remove_lmb(struct drmem_lmb *lmb)
+static int dlpar_remove_lmb(struct drmem_lmb *lmb,
+		enum dt_update_status *dt_update)
 {
 	unsigned long block_sz;
 	phys_addr_t base_addr;
-	int rc, nid;
+	int rc, ret, nid;
 
 	if (!lmb_is_removable(lmb))
 		return -EINVAL;

@@ -372,6 +381,13 @@ static int dlpar_remove_lmb(struct drmem_lmb *lmb)
 	invalidate_lmb_associativity_index(lmb);
 	lmb_clear_nid(lmb);
 	lmb->flags &= ~DRCONF_MEM_ASSIGNED;
+	if (*dt_update) {
+		ret = drmem_update_dt();
+		if (ret)
+			pr_warn("%s fail to update dt, but continue\n", __func__);
+		else
+			*dt_update = DT_UPDATED;
+	}
 
 	__remove_memory(nid, base_addr, block_sz);
 

@@ -387,6 +403,7 @@ static int dlpar_memory_remove_by_count(u32 lmbs_to_remove)
 	int lmbs_removed = 0;
 	int lmbs_available = 0;
 	int rc;
+	enum dt_update_status dt_update = DT_NOUPDATE;
 
 	pr_info("Attempting to hot-remove %d LMB(s)\n", lmbs_to_remove);
 

@@ -409,7 +426,11 @@ static int dlpar_memory_remove_by_count(u32 lmbs_to_remove)
 	}
 
 	for_each_drmem_lmb(lmb) {
-		rc = dlpar_remove_lmb(lmb);
+
+		/* combine dt updating for all LMBs */
+		if (lmbs_to_remove - lmbs_removed <= 1)
+			dt_update = DT_TOUPDATE;
+		rc = dlpar_remove_lmb(lmb, &dt_update);
 		if (rc)
 			continue;
 

@@ -424,13 +445,17 @@ static int dlpar_memory_remove_by_count(u32 lmbs_to_remove)
 	}
 
 	if (lmbs_removed != lmbs_to_remove) {
+		enum dt_update_status rollback_dt_update = DT_NOUPDATE;
+
 		pr_err("Memory hot-remove failed, adding LMB's back\n");
 
 		for_each_drmem_lmb(lmb) {
 			if (!drmem_lmb_reserved(lmb))
 				continue;
 
-			rc = dlpar_add_lmb(lmb);
+			if (--lmbs_removed == 0 && dt_update == DT_UPDATED)
+				rollback_dt_update = DT_TOUPDATE;
+			rc = dlpar_add_lmb(lmb, &rollback_dt_update);
 			if (rc)
 				pr_err("Failed to add LMB back, drc index %x\n",
 				       lmb->drc_index);

@@ -458,6 +483,7 @@ static int dlpar_memory_remove_by_count(u32 lmbs_to_remove)
 
 static int dlpar_memory_remove_by_index(u32 drc_index)
 {
+	enum dt_update_status dt_update = DT_TOUPDATE;
 	struct drmem_lmb *lmb;
 	int lmb_found;
 	int rc;

@@ -468,7 +494,7 @@ static int dlpar_memory_remove_by_index(u32 drc_index)
 	for_each_drmem_lmb(lmb) {
 		if (lmb->drc_index == drc_index) {
 			lmb_found = 1;
-			rc = dlpar_remove_lmb(lmb);
+			rc = dlpar_remove_lmb(lmb, &dt_update);
 			if (!rc)
 				dlpar_release_drc(lmb->drc_index);
 

@@ -490,6 +516,7 @@ static int dlpar_memory_remove_by_index(u32 drc_index)
 
 static int dlpar_memory_remove_by_ic(u32 lmbs_to_remove, u32 drc_index)
 {
+	enum dt_update_status dt_update = DT_NOUPDATE;
 	struct drmem_lmb *lmb, *start_lmb, *end_lmb;
 	int lmbs_available = 0;
 	int rc;

@@ -519,7 +546,9 @@ static int dlpar_memory_remove_by_ic(u32 lmbs_to_remove, u32 drc_index)
 		if (!(lmb->flags & DRCONF_MEM_ASSIGNED))
 			continue;
 
-		rc = dlpar_remove_lmb(lmb);
+		if (lmb == end_lmb)
+			dt_update = DT_TOUPDATE;
+		rc = dlpar_remove_lmb(lmb, &dt_update);
 		if (rc)
 			break;
 

@@ -527,14 +556,16 @@ static int dlpar_memory_remove_by_ic(u32 lmbs_to_remove, u32 drc_index)
 	}
 
 	if (rc) {
-		pr_err("Memory indexed-count-remove failed, adding any removed LMBs\n");
+		enum dt_update_status rollback_dt_update = DT_NOUPDATE;
 
+		pr_err("Memory indexed-count-remove failed, adding any removed LMBs\n");
 
 		for_each_drmem_lmb_in_range(lmb, start_lmb, end_lmb) {
 			if (!drmem_lmb_reserved(lmb))
 				continue;
-
-			rc = dlpar_add_lmb(lmb);
+			if (lmb == end_lmb && dt_update == DT_UPDATED)
+				rollback_dt_update = DT_TOUPDATE;
+			rc = dlpar_add_lmb(lmb, &rollback_dt_update);
 			if (rc)
 				pr_err("Failed to add LMB, drc index %x\n",
 				       lmb->drc_index);

@@ -572,7 +603,7 @@ static inline int dlpar_memory_remove(struct pseries_hp_errorlog *hp_elog)
 {
 	return -EOPNOTSUPP;
 }
-static int dlpar_remove_lmb(struct drmem_lmb *lmb)
+static int dlpar_remove_lmb(struct drmem_lmb *lmb, bool dt_update)
 {
 	return -EOPNOTSUPP;
 }

@@ -591,10 +622,11 @@ static int dlpar_memory_remove_by_ic(u32 lmbs_to_remove, u32 drc_index)
 }
 #endif /* CONFIG_MEMORY_HOTREMOVE */
 
-static int dlpar_add_lmb(struct drmem_lmb *lmb)
+static int dlpar_add_lmb(struct drmem_lmb *lmb,
+		enum dt_update_status *dt_update)
 {
 	unsigned long block_sz;
-	int rc;
+	int rc, ret;
 
 	if (lmb->flags & DRCONF_MEM_ASSIGNED)
 		return -EINVAL;

@@ -607,6 +639,13 @@ static int dlpar_add_lmb(struct drmem_lmb *lmb)
 
 	lmb_set_nid(lmb);
 	lmb->flags |= DRCONF_MEM_ASSIGNED;
+	if (*dt_update) {
+		ret = drmem_update_dt();
+		if (ret)
+			pr_warn("%s fail to update dt, but continue\n", __func__);
+		else
+			*dt_update = DT_UPDATED;
+	}
 
 	block_sz = memory_block_size_bytes();
 

@@ -616,6 +655,8 @@ static int dlpar_add_lmb(struct drmem_lmb *lmb)
 		invalidate_lmb_associativity_index(lmb);
 		lmb_clear_nid(lmb);
 		lmb->flags &= ~DRCONF_MEM_ASSIGNED;
+		if (*dt_update == DT_UPDATED)
+			drmem_update_dt();
 		return rc;
 	}
 

@@ -627,7 +668,11 @@ static int dlpar_add_lmb(struct drmem_lmb *lmb)
 		invalidate_lmb_associativity_index(lmb);
 		lmb_clear_nid(lmb);
 		lmb->flags &= ~DRCONF_MEM_ASSIGNED;
-
+		if (*dt_update == DT_UPDATED) {
+			ret = drmem_update_dt();
+			if (ret)
+				pr_warn("%s fail to update dt during rollback, but continue\n", __func__);
+		}
 		__remove_memory(nid, base_addr, block_sz);
 	}
 

@@ -636,6 +681,7 @@ static int dlpar_add_lmb(struct drmem_lmb *lmb)
 
 static int dlpar_memory_add_by_count(u32 lmbs_to_add)
 {
+	enum dt_update_status dt_update = DT_NOUPDATE;
 	struct drmem_lmb *lmb;
 	int lmbs_available = 0;
 	int lmbs_added = 0;

@@ -666,7 +712,9 @@ static int dlpar_memory_add_by_count(u32 lmbs_to_add)
 		if (rc)
 			continue;
 
-		rc = dlpar_add_lmb(lmb);
+		if (lmbs_to_add - lmbs_added <= 1)
+			dt_update = DT_TOUPDATE;
+		rc = dlpar_add_lmb(lmb, &dt_update);
 		if (rc) {
 			dlpar_release_drc(lmb->drc_index);
 			continue;

@@ -683,13 +731,18 @@ static int dlpar_memory_add_by_count(u32 lmbs_to_add)
 	}
 
 	if (lmbs_added != lmbs_to_add) {
+		enum dt_update_status rollback_dt_update = DT_NOUPDATE;
+
 		pr_err("Memory hot-add failed, removing any added LMBs\n");
 
 		for_each_drmem_lmb(lmb) {
 			if (!drmem_lmb_reserved(lmb))
 				continue;
 
-			rc = dlpar_remove_lmb(lmb);
+			if (--lmbs_added == 0 && dt_update == DT_UPDATED)
+				rollback_dt_update = DT_TOUPDATE;
+
+			rc = dlpar_remove_lmb(lmb, &rollback_dt_update);
 			if (rc)
 				pr_err("Failed to remove LMB, drc index %x\n",
 				       lmb->drc_index);

@@ -716,6 +769,7 @@ static int dlpar_memory_add_by_count(u32 lmbs_to_add)
 
 static int dlpar_memory_add_by_index(u32 drc_index)
 {
+	enum dt_update_status dt_update = DT_TOUPDATE;
 	struct drmem_lmb *lmb;
 	int rc, lmb_found;
 

@@ -727,7 +781,7 @@ static int dlpar_memory_add_by_index(u32 drc_index)
 			lmb_found = 1;
 			rc = dlpar_acquire_drc(lmb->drc_index);
 			if (!rc) {
-				rc = dlpar_add_lmb(lmb);
+				rc = dlpar_add_lmb(lmb, &dt_update);
 				if (rc)
 					dlpar_release_drc(lmb->drc_index);
 			}

@@ -750,6 +804,7 @@ static int dlpar_memory_add_by_index(u32 drc_index)
 
 static int dlpar_memory_add_by_ic(u32 lmbs_to_add, u32 drc_index)
 {
+	enum dt_update_status dt_update = DT_NOUPDATE;
 	struct drmem_lmb *lmb, *start_lmb, *end_lmb;
 	int lmbs_available = 0;
 	int rc;

@@ -783,7 +838,9 @@ static int dlpar_memory_add_by_ic(u32 lmbs_to_add, u32 drc_index)
 		if (rc)
 			break;
 
-		rc = dlpar_add_lmb(lmb);
+		if (lmb == end_lmb)
+			dt_update = DT_TOUPDATE;
+		rc = dlpar_add_lmb(lmb, &dt_update);
 		if (rc) {
 			dlpar_release_drc(lmb->drc_index);
 			break;

@@ -796,10 +853,14 @@ static int dlpar_memory_add_by_ic(u32 lmbs_to_add, u32 drc_index)
 		pr_err("Memory indexed-count-add failed, removing any added LMBs\n");
 
 		for_each_drmem_lmb_in_range(lmb, start_lmb, end_lmb) {
+			enum dt_update_status rollback_dt_update = DT_NOUPDATE;
+
 			if (!drmem_lmb_reserved(lmb))
 				continue;
 
-			rc = dlpar_remove_lmb(lmb);
+			if (lmb == end_lmb && dt_update == DT_UPDATED)
+				rollback_dt_update = DT_TOUPDATE;
+			rc = dlpar_remove_lmb(lmb, &rollback_dt_update);
 			if (rc)
 				pr_err("Failed to remove LMB, drc index %x\n",
 				       lmb->drc_index);

@@ -879,9 +940,6 @@ int dlpar_memory(struct pseries_hp_errorlog *hp_elog)
 		break;
 	}
 
-	if (!rc)
-		rc = drmem_update_dt();
-
 	unlock_device_hotplug();
 	return rc;
 }

-- 
2.7.5

Hello guys. Do you have further comments on this version?

Thanks,
Pingfan

On Mon, Aug 10, 2020 at 4:53 PM Pingfan Liu [off-list ref] wrote:

Le 10/08/2020 à 10:52, Pingfan Liu a écrit :

That test is wrong, you should do:
         if (*dt_update && *dt_update == DT_TOUPDATE) {

With the current code, the device tree is updated all the time.

Another option would be to pass a valid pointer (!= NULL) only when DT update is 
required, this way you don't need the DT_TOUPDATE value. The caller would have 
to set the pointer accordingly. The advantage with this option is the caller is 
guaranteed that its variable is not touched by the callee when no device tree is 
requested. A simple boolean pointer would be enough without the need to this enum.

Le 10/08/2020 à 10:52, Pingfan Liu a écrit :

This patch prepares for the incoming patch which swaps the order of
KOBJ_ADD/REMOVE uevent and dt's updating.

The dt updating should come after lmb operations, and before
__remove_memory()/__add_memory().  Accordingly, grouping all lmb operations
before the memblock's.

I can't find the link between this commit description and the code's changes below.

On Thu, Aug 27, 2020 at 3:53 PM Laurent Dufour [off-list ref] wrote:

Le 10/08/2020 à 10:52, Pingfan Liu a écrit :

quoted

A bug is observed on pseries by taking the following steps on rhel:
-1. drmgr -c mem -r -q 5
-2. echo c > /proc/sysrq-trigger

And then, the failure looks like:
kdump: saving to /sysroot//var/crash/127.0.0.1-2020-01-16-02:06:14/
kdump: saving vmcore-dmesg.txt
kdump: saving vmcore-dmesg.txt complete
kdump: saving vmcore
  Checking for memory holes                         : [  0.0 %] /                   Checking for memory holes                         : [100.0 %] |                   Excluding unnecessary pages                       : [100.0 %] \                   Copying data                                      : [  0.3 %] -          eta: 38s[   44.337636] hash-mmu: mm: Hashing failure ! EA=0x7fffba400000 access=0x8000000000000004 current=makedumpfile
[   44.337663] hash-mmu:     trap=0x300 vsid=0x13a109c ssize=1 base psize=2 psize 2 pte=0xc000000050000504
[   44.337677] hash-mmu: mm: Hashing failure ! EA=0x7fffba400000 access=0x8000000000000004 current=makedumpfile
[   44.337692] hash-mmu:     trap=0x300 vsid=0x13a109c ssize=1 base psize=2 psize 2 pte=0xc000000050000504
[   44.337708] makedumpfile[469]: unhandled signal 7 at 00007fffba400000 nip 00007fffbbc4d7fc lr 000000011356ca3c code 2
[   44.338548] Core dump to |/bin/false pipe failed
/lib/kdump-lib-initramfs.sh: line 98:   469 Bus error               $CORE_COLLECTOR /proc/vmcore $_mp/$KDUMP_PATH/$HOST_IP-$DATEDIR/vmcore-incomplete
kdump: saving vmcore failed

* Root cause *
   After analyzing, it turns out that in the current implementation,
when hot-removing lmb, the KOBJ_REMOVE event ejects before the dt updating as
the code __remove_memory() comes before drmem_update_dt().
So in kdump kernel, when read_from_oldmem() resorts to
pSeries_lpar_hpte_insert() to install hpte, but fails with -2 due to
non-exist pfn. And finally, low_hash_fault() raise SIGBUS to process, as it
can be observed "Bus error"

 From a viewpoint of listener and publisher, the publisher notifies the
listener before data is ready.  This introduces a problem where udev
launches kexec-tools (due to KOBJ_REMOVE) and loads a stale dt before
updating. And in capture kernel, makedumpfile will access the memory based
on the stale dt info, and hit a SIGBUS error due to an un-existed lmb.

* Fix *
This bug is introduced by commit 063b8b1251fd
("powerpc/pseries/memory-hotplug: Only update DT once per memory DLPAR
request"), which tried to combine all the dt updating into one.

To fix this issue, meanwhile not to introduce a quadratic runtime
complexity by the model:
   dlpar_memory_add_by_count
     for_each_drmem_lmb             <--
       dlpar_add_lmb
         drmem_update_dt(_v1|_v2)
           for_each_drmem_lmb       <--
The dt should still be only updated once, and just before the last memory
online/offline event is ejected to user space. Achieve this by tracing the
num of lmb added or removed.

Signed-off-by: Pingfan Liu <redacted>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Hari Bathini <hbathini@linux.ibm.com>
Cc: Nathan Lynch <redacted>
Cc: Nathan Fontenot <redacted>
Cc: Laurent Dufour <redacted>
To: linuxppc-dev@lists.ozlabs.org
Cc: kexec@lists.infradead.org
---
v4 -> v5: change dlpar_add_lmb()/dlpar_remove_lmb() prototype to report
           whether dt is updated successfully.
           Fix a condition boundary check bug
v3 -> v4: resolve a quadratic runtime complexity issue.
           This series is applied on next-test branch
  arch/powerpc/platforms/pseries/hotplug-memory.c | 102 +++++++++++++++++++-----
  1 file changed, 80 insertions(+), 22 deletions(-)

diff --git a/arch/powerpc/platforms/pseries/hotplug-memory.c b/arch/powerpc/platforms/pseries/hotplug-memory.c
index 46cbcd1..1567d9f 100644
--- a/arch/powerpc/platforms/pseries/hotplug-memory.c
+++ b/arch/powerpc/platforms/pseries/hotplug-memory.c

@@ -350,13 +350,22 @@ static bool lmb_is_removable(struct drmem_lmb *lmb)
      return true;
  }

-static int dlpar_add_lmb(struct drmem_lmb *);
+enum dt_update_status {
+     DT_NOUPDATE,
+     DT_TOUPDATE,
+     DT_UPDATED,
+};
+
+/* "*dt_update" returns DT_UPDATED if updated */
+static int dlpar_add_lmb(struct drmem_lmb *lmb,
+             enum dt_update_status *dt_update);

-static int dlpar_remove_lmb(struct drmem_lmb *lmb)
+static int dlpar_remove_lmb(struct drmem_lmb *lmb,
+             enum dt_update_status *dt_update)
  {
      unsigned long block_sz;
      phys_addr_t base_addr;
-     int rc, nid;
+     int rc, ret, nid;

      if (!lmb_is_removable(lmb))
              return -EINVAL;

@@ -372,6 +381,13 @@ static int dlpar_remove_lmb(struct drmem_lmb *lmb)
      invalidate_lmb_associativity_index(lmb);
      lmb_clear_nid(lmb);
      lmb->flags &= ~DRCONF_MEM_ASSIGNED;
+     if (*dt_update) {

Original, I plan to use it to exclude the case of DT_NOUPDATE, which is value 0.
And I think it looks better by using if (*dt_update == DT_TOUPDATE)

That test is wrong, you should do:
         if (*dt_update && *dt_update == DT_TOUPDATE) {

I think you mean  if (dt_update && *dt_update == DT_TOUPDATE) {

With the current code, the device tree is updated all the time.

Another option would be to pass a valid pointer (!= NULL) only when DT update is
required, this way you don't need the DT_TOUPDATE value. The caller would have
to set the pointer accordingly. The advantage with this option is the caller is
guaranteed that its variable is not touched by the callee when no device tree is
requested. A simple boolean pointer would be enough without the need to this enum.

It is expected that dlpar_remove_lmb/dlpar_add_lmb can report whether
they successfully update dt or not. So the caller can handle the
different cases.

Thanks,
Pingfan

Hello,

On Fri, Aug 28, 2020 at 04:10:09PM +0800, Pingfan Liu wrote:

On Thu, Aug 27, 2020 at 3:53 PM Laurent Dufour [off-list ref] wrote:

quoted

Le 10/08/2020 à 10:52, Pingfan Liu a écrit :

quoted

A bug is observed on pseries by taking the following steps on rhel:
-1. drmgr -c mem -r -q 5
-2. echo c > /proc/sysrq-trigger

And then, the failure looks like:
kdump: saving to /sysroot//var/crash/127.0.0.1-2020-01-16-02:06:14/
kdump: saving vmcore-dmesg.txt
kdump: saving vmcore-dmesg.txt complete
kdump: saving vmcore
  Checking for memory holes                         : [  0.0 %] /                   Checking for memory holes                         : [100.0 %] |                   Excluding unnecessary pages                       : [100.0 %] \                   Copying data                                      : [  0.3 %] -          eta: 38s[   44.337636] hash-mmu: mm: Hashing failure ! EA=0x7fffba400000 access=0x8000000000000004 current=makedumpfile
[   44.337663] hash-mmu:     trap=0x300 vsid=0x13a109c ssize=1 base psize=2 psize 2 pte=0xc000000050000504
[   44.337677] hash-mmu: mm: Hashing failure ! EA=0x7fffba400000 access=0x8000000000000004 current=makedumpfile
[   44.337692] hash-mmu:     trap=0x300 vsid=0x13a109c ssize=1 base psize=2 psize 2 pte=0xc000000050000504
[   44.337708] makedumpfile[469]: unhandled signal 7 at 00007fffba400000 nip 00007fffbbc4d7fc lr 000000011356ca3c code 2
[   44.338548] Core dump to |/bin/false pipe failed
/lib/kdump-lib-initramfs.sh: line 98:   469 Bus error               $CORE_COLLECTOR /proc/vmcore $_mp/$KDUMP_PATH/$HOST_IP-$DATEDIR/vmcore-incomplete
kdump: saving vmcore failed

* Root cause *
   After analyzing, it turns out that in the current implementation,
when hot-removing lmb, the KOBJ_REMOVE event ejects before the dt updating as
the code __remove_memory() comes before drmem_update_dt().
So in kdump kernel, when read_from_oldmem() resorts to
pSeries_lpar_hpte_insert() to install hpte, but fails with -2 due to
non-exist pfn. And finally, low_hash_fault() raise SIGBUS to process, as it
can be observed "Bus error"

 From a viewpoint of listener and publisher, the publisher notifies the
listener before data is ready.  This introduces a problem where udev
launches kexec-tools (due to KOBJ_REMOVE) and loads a stale dt before
updating. And in capture kernel, makedumpfile will access the memory based
on the stale dt info, and hit a SIGBUS error due to an un-existed lmb.

* Fix *
This bug is introduced by commit 063b8b1251fd
("powerpc/pseries/memory-hotplug: Only update DT once per memory DLPAR
request"), which tried to combine all the dt updating into one.

To fix this issue, meanwhile not to introduce a quadratic runtime
complexity by the model:
   dlpar_memory_add_by_count
     for_each_drmem_lmb             <--
       dlpar_add_lmb
         drmem_update_dt(_v1|_v2)
           for_each_drmem_lmb       <--
The dt should still be only updated once, and just before the last memory
online/offline event is ejected to user space. Achieve this by tracing the
num of lmb added or removed.

Signed-off-by: Pingfan Liu <redacted>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Hari Bathini <hbathini@linux.ibm.com>
Cc: Nathan Lynch <redacted>
Cc: Nathan Fontenot <redacted>
Cc: Laurent Dufour <redacted>
To: linuxppc-dev@lists.ozlabs.org
Cc: kexec@lists.infradead.org
---
v4 -> v5: change dlpar_add_lmb()/dlpar_remove_lmb() prototype to report
           whether dt is updated successfully.
           Fix a condition boundary check bug
v3 -> v4: resolve a quadratic runtime complexity issue.
           This series is applied on next-test branch
  arch/powerpc/platforms/pseries/hotplug-memory.c | 102 +++++++++++++++++++-----
  1 file changed, 80 insertions(+), 22 deletions(-)

diff --git a/arch/powerpc/platforms/pseries/hotplug-memory.c b/arch/powerpc/platforms/pseries/hotplug-memory.c
index 46cbcd1..1567d9f 100644
--- a/arch/powerpc/platforms/pseries/hotplug-memory.c
+++ b/arch/powerpc/platforms/pseries/hotplug-memory.c

@@ -350,13 +350,22 @@ static bool lmb_is_removable(struct drmem_lmb *lmb)
      return true;
  }

-static int dlpar_add_lmb(struct drmem_lmb *);
+enum dt_update_status {
+     DT_NOUPDATE,
+     DT_TOUPDATE,
+     DT_UPDATED,
+};
+
+/* "*dt_update" returns DT_UPDATED if updated */
+static int dlpar_add_lmb(struct drmem_lmb *lmb,
+             enum dt_update_status *dt_update);

-static int dlpar_remove_lmb(struct drmem_lmb *lmb)
+static int dlpar_remove_lmb(struct drmem_lmb *lmb,
+             enum dt_update_status *dt_update)
  {
      unsigned long block_sz;
      phys_addr_t base_addr;
-     int rc, nid;
+     int rc, ret, nid;

      if (!lmb_is_removable(lmb))
              return -EINVAL;

@@ -372,6 +381,13 @@ static int dlpar_remove_lmb(struct drmem_lmb *lmb)
      invalidate_lmb_associativity_index(lmb);
      lmb_clear_nid(lmb);
      lmb->flags &= ~DRCONF_MEM_ASSIGNED;
+     if (*dt_update) {

Original, I plan to use it to exclude the case of DT_NOUPDATE, which is value 0.
And I think it looks better by using if (*dt_update == DT_TOUPDATE)

quoted

That test is wrong, you should do:
         if (*dt_update && *dt_update == DT_TOUPDATE) {

I think you mean  if (dt_update && *dt_update == DT_TOUPDATE) {

quoted

With the current code, the device tree is updated all the time.

Another option would be to pass a valid pointer (!= NULL) only when DT update is
required, this way you don't need the DT_TOUPDATE value. The caller would have
to set the pointer accordingly. The advantage with this option is the caller is
guaranteed that its variable is not touched by the callee when no device tree is
requested. A simple boolean pointer would be enough without the need to this enum.

It is expected that dlpar_remove_lmb/dlpar_add_lmb can report whether
they successfully update dt or not. So the caller can handle the
different cases.

Is there any plan to refresh this patch to apply to master?

I am using an older revision of this patch so I am not in the position
to repost an updated version.

I lack some otimization in my patch so I probably have the quadratic
coplexity of the add mentioned above.

Thanks

Michal

On Sat, Apr 10, 2021 at 12:33 AM Michal Suchánek [off-list ref] wrote:

Hello,

On Fri, Aug 28, 2020 at 04:10:09PM +0800, Pingfan Liu wrote:

quoted

On Thu, Aug 27, 2020 at 3:53 PM Laurent Dufour [off-list ref] wrote:

quoted

Le 10/08/2020 à 10:52, Pingfan Liu a écrit :

quoted

A bug is observed on pseries by taking the following steps on rhel:
-1. drmgr -c mem -r -q 5
-2. echo c > /proc/sysrq-trigger

And then, the failure looks like:
kdump: saving to /sysroot//var/crash/127.0.0.1-2020-01-16-02:06:14/
kdump: saving vmcore-dmesg.txt
kdump: saving vmcore-dmesg.txt complete
kdump: saving vmcore
  Checking for memory holes                         : [  0.0 %] /                   Checking for memory holes                         : [100.0 %] |                   Excluding unnecessary pages                       : [100.0 %] \                   Copying data                                      : [  0.3 %] -          eta: 38s[   44.337636] hash-mmu: mm: Hashing failure ! EA=0x7fffba400000 access=0x8000000000000004 current=makedumpfile
[   44.337663] hash-mmu:     trap=0x300 vsid=0x13a109c ssize=1 base psize=2 psize 2 pte=0xc000000050000504
[   44.337677] hash-mmu: mm: Hashing failure ! EA=0x7fffba400000 access=0x8000000000000004 current=makedumpfile
[   44.337692] hash-mmu:     trap=0x300 vsid=0x13a109c ssize=1 base psize=2 psize 2 pte=0xc000000050000504
[   44.337708] makedumpfile[469]: unhandled signal 7 at 00007fffba400000 nip 00007fffbbc4d7fc lr 000000011356ca3c code 2
[   44.338548] Core dump to |/bin/false pipe failed
/lib/kdump-lib-initramfs.sh: line 98:   469 Bus error               $CORE_COLLECTOR /proc/vmcore $_mp/$KDUMP_PATH/$HOST_IP-$DATEDIR/vmcore-incomplete
kdump: saving vmcore failed

* Root cause *
   After analyzing, it turns out that in the current implementation,
when hot-removing lmb, the KOBJ_REMOVE event ejects before the dt updating as
the code __remove_memory() comes before drmem_update_dt().
So in kdump kernel, when read_from_oldmem() resorts to
pSeries_lpar_hpte_insert() to install hpte, but fails with -2 due to
non-exist pfn. And finally, low_hash_fault() raise SIGBUS to process, as it
can be observed "Bus error"

 From a viewpoint of listener and publisher, the publisher notifies the
listener before data is ready.  This introduces a problem where udev
launches kexec-tools (due to KOBJ_REMOVE) and loads a stale dt before
updating. And in capture kernel, makedumpfile will access the memory based
on the stale dt info, and hit a SIGBUS error due to an un-existed lmb.

* Fix *
This bug is introduced by commit 063b8b1251fd
("powerpc/pseries/memory-hotplug: Only update DT once per memory DLPAR
request"), which tried to combine all the dt updating into one.

To fix this issue, meanwhile not to introduce a quadratic runtime
complexity by the model:
   dlpar_memory_add_by_count
     for_each_drmem_lmb             <--
       dlpar_add_lmb
         drmem_update_dt(_v1|_v2)
           for_each_drmem_lmb       <--
The dt should still be only updated once, and just before the last memory
online/offline event is ejected to user space. Achieve this by tracing the
num of lmb added or removed.

Signed-off-by: Pingfan Liu <redacted>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Hari Bathini <hbathini@linux.ibm.com>
Cc: Nathan Lynch <redacted>
Cc: Nathan Fontenot <redacted>
Cc: Laurent Dufour <redacted>
To: linuxppc-dev@lists.ozlabs.org
Cc: kexec@lists.infradead.org
---
v4 -> v5: change dlpar_add_lmb()/dlpar_remove_lmb() prototype to report
           whether dt is updated successfully.
           Fix a condition boundary check bug
v3 -> v4: resolve a quadratic runtime complexity issue.
           This series is applied on next-test branch
  arch/powerpc/platforms/pseries/hotplug-memory.c | 102 +++++++++++++++++++-----
  1 file changed, 80 insertions(+), 22 deletions(-)

diff --git a/arch/powerpc/platforms/pseries/hotplug-memory.c b/arch/powerpc/platforms/pseries/hotplug-memory.c
index 46cbcd1..1567d9f 100644
--- a/arch/powerpc/platforms/pseries/hotplug-memory.c
+++ b/arch/powerpc/platforms/pseries/hotplug-memory.c

@@ -350,13 +350,22 @@ static bool lmb_is_removable(struct drmem_lmb *lmb)
      return true;
  }

-static int dlpar_add_lmb(struct drmem_lmb *);
+enum dt_update_status {
+     DT_NOUPDATE,
+     DT_TOUPDATE,
+     DT_UPDATED,
+};
+
+/* "*dt_update" returns DT_UPDATED if updated */
+static int dlpar_add_lmb(struct drmem_lmb *lmb,
+             enum dt_update_status *dt_update);

-static int dlpar_remove_lmb(struct drmem_lmb *lmb)
+static int dlpar_remove_lmb(struct drmem_lmb *lmb,
+             enum dt_update_status *dt_update)
  {
      unsigned long block_sz;
      phys_addr_t base_addr;
-     int rc, nid;
+     int rc, ret, nid;

      if (!lmb_is_removable(lmb))
              return -EINVAL;

@@ -372,6 +381,13 @@ static int dlpar_remove_lmb(struct drmem_lmb *lmb)
      invalidate_lmb_associativity_index(lmb);
      lmb_clear_nid(lmb);
      lmb->flags &= ~DRCONF_MEM_ASSIGNED;
+     if (*dt_update) {

Original, I plan to use it to exclude the case of DT_NOUPDATE, which is value 0.
And I think it looks better by using if (*dt_update == DT_TOUPDATE)

quoted

That test is wrong, you should do:
         if (*dt_update && *dt_update == DT_TOUPDATE) {

I think you mean  if (dt_update && *dt_update == DT_TOUPDATE) {

quoted

With the current code, the device tree is updated all the time.

Another option would be to pass a valid pointer (!= NULL) only when DT update is
required, this way you don't need the DT_TOUPDATE value. The caller would have
to set the pointer accordingly. The advantage with this option is the caller is
guaranteed that its variable is not touched by the callee when no device tree is
requested. A simple boolean pointer would be enough without the need to this enum.

It is expected that dlpar_remove_lmb/dlpar_add_lmb can report whether
they successfully update dt or not. So the caller can handle the
different cases.

Is there any plan to refresh this patch to apply to master?

Spreading the failure and rollback logic around the code looks urgly.
Even myself dislike it. But I have not found a better way out.

I am using an older revision of this patch so I am not in the position
to repost an updated version.

I lack some otimization in my patch so I probably have the quadratic
coplexity of the add mentioned above.

Do you have any good ideas?

Thanks,
Pingfan

Hello,

On Wed, Apr 14, 2021 at 11:08:19AM +0800, Pingfan Liu wrote:

On Sat, Apr 10, 2021 at 12:33 AM Michal Suchánek [off-list ref] wrote:

quoted

Hello,

On Fri, Aug 28, 2020 at 04:10:09PM +0800, Pingfan Liu wrote:

quoted

On Thu, Aug 27, 2020 at 3:53 PM Laurent Dufour [off-list ref] wrote:

quoted

Le 10/08/2020 à 10:52, Pingfan Liu a écrit :

quoted

A bug is observed on pseries by taking the following steps on rhel:
-1. drmgr -c mem -r -q 5
-2. echo c > /proc/sysrq-trigger

And then, the failure looks like:
kdump: saving to /sysroot//var/crash/127.0.0.1-2020-01-16-02:06:14/
kdump: saving vmcore-dmesg.txt
kdump: saving vmcore-dmesg.txt complete
kdump: saving vmcore
  Checking for memory holes                         : [  0.0 %] /                   Checking for memory holes                         : [100.0 %] |                   Excluding unnecessary pages                       : [100.0 %] \                   Copying data                                      : [  0.3 %] -          eta: 38s[   44.337636] hash-mmu: mm: Hashing failure ! EA=0x7fffba400000 access=0x8000000000000004 current=makedumpfile
[   44.337663] hash-mmu:     trap=0x300 vsid=0x13a109c ssize=1 base psize=2 psize 2 pte=0xc000000050000504
[   44.337677] hash-mmu: mm: Hashing failure ! EA=0x7fffba400000 access=0x8000000000000004 current=makedumpfile
[   44.337692] hash-mmu:     trap=0x300 vsid=0x13a109c ssize=1 base psize=2 psize 2 pte=0xc000000050000504
[   44.337708] makedumpfile[469]: unhandled signal 7 at 00007fffba400000 nip 00007fffbbc4d7fc lr 000000011356ca3c code 2
[   44.338548] Core dump to |/bin/false pipe failed
/lib/kdump-lib-initramfs.sh: line 98:   469 Bus error               $CORE_COLLECTOR /proc/vmcore $_mp/$KDUMP_PATH/$HOST_IP-$DATEDIR/vmcore-incomplete
kdump: saving vmcore failed

* Root cause *
   After analyzing, it turns out that in the current implementation,
when hot-removing lmb, the KOBJ_REMOVE event ejects before the dt updating as
the code __remove_memory() comes before drmem_update_dt().
So in kdump kernel, when read_from_oldmem() resorts to
pSeries_lpar_hpte_insert() to install hpte, but fails with -2 due to
non-exist pfn. And finally, low_hash_fault() raise SIGBUS to process, as it
can be observed "Bus error"

 From a viewpoint of listener and publisher, the publisher notifies the
listener before data is ready.  This introduces a problem where udev
launches kexec-tools (due to KOBJ_REMOVE) and loads a stale dt before
updating. And in capture kernel, makedumpfile will access the memory based
on the stale dt info, and hit a SIGBUS error due to an un-existed lmb.

* Fix *
This bug is introduced by commit 063b8b1251fd
("powerpc/pseries/memory-hotplug: Only update DT once per memory DLPAR
request"), which tried to combine all the dt updating into one.

To fix this issue, meanwhile not to introduce a quadratic runtime
complexity by the model:
   dlpar_memory_add_by_count
     for_each_drmem_lmb             <--
       dlpar_add_lmb
         drmem_update_dt(_v1|_v2)
           for_each_drmem_lmb       <--
The dt should still be only updated once, and just before the last memory
online/offline event is ejected to user space. Achieve this by tracing the
num of lmb added or removed.

Signed-off-by: Pingfan Liu <redacted>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Hari Bathini <hbathini@linux.ibm.com>
Cc: Nathan Lynch <redacted>
Cc: Nathan Fontenot <redacted>
Cc: Laurent Dufour <redacted>
To: linuxppc-dev@lists.ozlabs.org
Cc: kexec@lists.infradead.org
---
v4 -> v5: change dlpar_add_lmb()/dlpar_remove_lmb() prototype to report
           whether dt is updated successfully.
           Fix a condition boundary check bug
v3 -> v4: resolve a quadratic runtime complexity issue.
           This series is applied on next-test branch
  arch/powerpc/platforms/pseries/hotplug-memory.c | 102 +++++++++++++++++++-----
  1 file changed, 80 insertions(+), 22 deletions(-)

diff --git a/arch/powerpc/platforms/pseries/hotplug-memory.c b/arch/powerpc/platforms/pseries/hotplug-memory.c
index 46cbcd1..1567d9f 100644
--- a/arch/powerpc/platforms/pseries/hotplug-memory.c
+++ b/arch/powerpc/platforms/pseries/hotplug-memory.c

@@ -350,13 +350,22 @@ static bool lmb_is_removable(struct drmem_lmb *lmb)
      return true;
  }

-static int dlpar_add_lmb(struct drmem_lmb *);
+enum dt_update_status {
+     DT_NOUPDATE,
+     DT_TOUPDATE,
+     DT_UPDATED,
+};
+
+/* "*dt_update" returns DT_UPDATED if updated */
+static int dlpar_add_lmb(struct drmem_lmb *lmb,
+             enum dt_update_status *dt_update);

-static int dlpar_remove_lmb(struct drmem_lmb *lmb)
+static int dlpar_remove_lmb(struct drmem_lmb *lmb,
+             enum dt_update_status *dt_update)
  {
      unsigned long block_sz;
      phys_addr_t base_addr;
-     int rc, nid;
+     int rc, ret, nid;

      if (!lmb_is_removable(lmb))
              return -EINVAL;

@@ -372,6 +381,13 @@ static int dlpar_remove_lmb(struct drmem_lmb *lmb)
      invalidate_lmb_associativity_index(lmb);
      lmb_clear_nid(lmb);
      lmb->flags &= ~DRCONF_MEM_ASSIGNED;
+     if (*dt_update) {

Original, I plan to use it to exclude the case of DT_NOUPDATE, which is value 0.
And I think it looks better by using if (*dt_update == DT_TOUPDATE)

quoted

That test is wrong, you should do:
         if (*dt_update && *dt_update == DT_TOUPDATE) {

I think you mean  if (dt_update && *dt_update == DT_TOUPDATE) {

quoted

With the current code, the device tree is updated all the time.

Another option would be to pass a valid pointer (!= NULL) only when DT update is
required, this way you don't need the DT_TOUPDATE value. The caller would have
to set the pointer accordingly. The advantage with this option is the caller is
guaranteed that its variable is not touched by the callee when no device tree is
requested. A simple boolean pointer would be enough without the need to this enum.

It is expected that dlpar_remove_lmb/dlpar_add_lmb can report whether
they successfully update dt or not. So the caller can handle the
different cases.

Is there any plan to refresh this patch to apply to master?

Spreading the failure and rollback logic around the code looks urgly.
Even myself dislike it. But I have not found a better way out.

quoted

I am using an older revision of this patch so I am not in the position
to repost an updated version.

I lack some otimization in my patch so I probably have the quadratic
coplexity of the add mentioned above.

Do you have any good ideas?

I am not very familiar with the code and I am not sure I will have time
to look closer at it anytime soon.

Computation complexity is traded for code complexity and it is not clear
if the tradeoff is good.

As a way forward I would suggest to split the patch into a minimal
correctness fix and an additional optimization that avoids needless
computation.

That way the extra code complexity and the speed improvement brought by
the optimization can be reviewed more easily.

I don't have myself access to any big machine so can't tell if the
optimization solves a real world problem either.

Thanks

Michal