Re: [PATCH v12 09/11] hmat acpi: Build System Locality Latency and Bandwidth Information Structure(s)

[Tao Xu]

On 10/11/2019 10:08 PM, Igor Mammedov wrote:
On Thu, 10 Oct 2019 14:53:56 +0800
Tao Xu <tao3...@intel.com> wrote:
On 10/3/2019 10:41 PM, Igor Mammedov wrote:
On Fri, 20 Sep 2019 15:43:47 +0800
Tao Xu <tao3...@intel.com> wrote:
   
From: Liu Jingqi <jingqi....@intel.com>

This structure describes the memory access latency and bandwidth
information from various memory access initiator proximity domains.
The latency and bandwidth numbers represented in this structure
correspond to rated latency and bandwidth for the platform.
The software could use this information as hint for optimization.

Signed-off-by: Liu Jingqi <jingqi....@intel.com>
Signed-off-by: Tao Xu <tao3...@intel.com>
---

Changes in v12:
      - Fix a bug that if HMAT is enabled and without hmat-lb setting,
        QEMU will crash. (reported by Danmei Wei)

Changes in v11:
      - Calculate base in build_hmat_lb().
---
   hw/acpi/hmat.c | 126 ++++++++++++++++++++++++++++++++++++++++++++++++-
   hw/acpi/hmat.h |   2 +
   2 files changed, 127 insertions(+), 1 deletion(-)

diff --git a/hw/acpi/hmat.c b/hw/acpi/hmat.c
index 1368fce7ee..e7be849581 100644
--- a/hw/acpi/hmat.c
+++ b/hw/acpi/hmat.c
@@ -27,6 +27,7 @@
   #include "qemu/osdep.h"
   #include "sysemu/numa.h"
   #include "hw/acpi/hmat.h"
+#include "qemu/error-report.h"
   
   /*
    * ACPI 6.3:
@@ -67,11 +68,105 @@ static void build_hmat_mpda(GArray *table_data, uint16_t 
flags, int initiator,
       build_append_int_noprefix(table_data, 0, 8);
   }
   
+static bool entry_overflow(uint64_t *lb_data, uint64_t base, int len)
+{
+    int i;
+
+    for (i = 0; i < len; i++) {
+        if (lb_data[i] / base >= UINT16_MAX) {
+            return true;
+        }
+    }
+
+    return false;
+}
I suggest to do this check at CLI parsing time
   
+/*
+ * ACPI 6.3: 5.2.27.4 System Locality Latency and Bandwidth Information
+ * Structure: Table 5-146
+ */
+static void build_hmat_lb(GArray *table_data, HMAT_LB_Info *hmat_lb,
+                          uint32_t num_initiator, uint32_t num_target,
+                          uint32_t *initiator_list, int type)
+{
+    uint8_t mask = 0x0f;
+    uint32_t s = num_initiator;
+    uint32_t t = num_target;
drop this locals and use arguments directly
   
+    uint64_t base = 1;
+    uint64_t *lb_data;
+    int i, unit;
+
+    /* Type */
+    build_append_int_noprefix(table_data, 1, 2);
+    /* Reserved */
+    build_append_int_noprefix(table_data, 0, 2);
+    /* Length */
+    build_append_int_noprefix(table_data, 32 + 4 * s + 4 * t + 2 * s * t, 4);
                                               ^^^^
to me above looks like /dev/random output, absolutely unreadable.
Suggest to use local var (like: lb_length) for expression with comments
beside magic numbers.
   
+    /* Flags: Bits [3:0] Memory Hierarchy, Bits[7:4] Reserved */
+    build_append_int_noprefix(table_data, hmat_lb->hierarchy & mask, 1);

why do you need to use mask here?
   
Because Bits[7:4] Reserved, so I use mask to keep it reserved.

these bits are not user provided and set to 0, if they get set it's
programming error and instead of masking problem out QEMU should abort,
I suggest replace masking with assert(!foo>>x).


+    /* Data Type */
+    build_append_int_noprefix(table_data, hmat_lb->data_type, 1);

Isn't hmat_lb->data_type and passed argument 'type' the same?
   
Yes, I will drop 'type'.
   
+    /* Reserved */
+    build_append_int_noprefix(table_data, 0, 2);
+    /* Number of Initiator Proximity Domains (s) */
+    build_append_int_noprefix(table_data, s, 4);
+    /* Number of Target Proximity Domains (t) */
+    build_append_int_noprefix(table_data, t, 4);
+    /* Reserved */
+    build_append_int_noprefix(table_data, 0, 4);
+
+    if (HMAT_IS_LATENCY(type)) {
+        unit = 1000;
+        lb_data = hmat_lb->latency;
+    } else {
+        unit = 1024;
+        lb_data = hmat_lb->bandwidth;
+    }
+
+    while (entry_overflow(lb_data, base, s * t)) {
+        for (i = 0; i < s * t; i++) {
+            if (!QEMU_IS_ALIGNED(lb_data[i], unit * base)) {
+                error_report("Invalid latency/bandwidth input, all "
+                "latencies/bandwidths should be specified in the same units.");
+                exit(1);
+            }
+        }
+        base *= unit;
+    }
Can you clarify what you are trying to check here?
   
This part I use entry_overflow() to check if uint16 can store entry. If
can't store and the entries matrix can be divisible by unit * base, then
base will be unit * base.

For example, if lb_data[i] are 1048576(1TB/s) and 1024(1GB/s), unit is
1024, so 1048576 is bigger than UINT16_MAX, and can be divisible by 1024
* 1, so base is 1024 and entries are 1024 and 1 (see entry =
hmat_lb->latency[i] / base;). The benefit is even user input different
unit(TB/s vs GB/s), we can still store the data as far as possible.

Is it possible instead of doing multiple iterations over lb_data
until it finds valid base, just go over lb_data once to find MIN/MAX
and then calculate base using it. Error out with max/min offending
values if it's not possible to compress the range into uint16_t?


Although we tell user input same unit data, such as use 1GB/s 3GB/s. If 
user input data such as 1048575, 1048576(1TB/s) and 1024(1GB/s), then we 
will get 1024 * (1023 1024 1). I am wondering if it is appropriate 
because we lose a float number(0.999020). But in our codes, it will 
raise error.