|
From: | Tao Xu |
Subject: | Re: [PATCH v12 09/11] hmat acpi: Build System Locality Latency and Bandwidth Information Structure(s) |
Date: | Tue, 15 Oct 2019 13:40:54 +0800 |
User-agent: | Mozilla/5.0 (Windows NT 10.0; WOW64; rv:60.0) Gecko/20100101 Thunderbird/60.9.0 |
On 10/15/2019 8:59 AM, Tao Xu wrote:
I understand the MIN/MAX means, in the case above, we get MAX is 9GB/s, MIN is 3GB/s, then I use code below to calculate :On 10/14/2019 5:00 PM, Igor Mammedov wrote:Sorry, I mean how we treat the data cannot be divisible if we use max/min as base. For another example, If user input the data(including 3 bandwidths) : 9GB/s 5GB/s 3GB/s. Then max/min result is 3. But entries should be uint16, (5GB/s)/3 we can only get 1GB/s, then we should raise error(overflow).On Sat, 12 Oct 2019 11:04:03 +0800 Tao Xu <address@hidden> wrote:On 10/11/2019 10:08 PM, Igor Mammedov wrote:On Thu, 10 Oct 2019 14:53:56 +0800 Tao Xu <address@hidden> wrote:On 10/3/2019 10:41 PM, Igor Mammedov wrote:On Fri, 20 Sep 2019 15:43:47 +0800 Tao Xu <address@hidden> wrote:From: Liu Jingqi <address@hidden> 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 <address@hidden> Signed-off-by: Tao Xu <address@hidden> --- 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 commentsbeside 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'.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+ /* 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?base will be unit * base. For example, if lb_data[i] are 1048576(1TB/s) and 1024(1GB/s), unit is1024, 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.I do not understand what you are trying to say here, could you rephrase it, so the problem would be more clear, please?But if this patch, we will get the base is 1GB/s.
while (max_data >= UINT16_MAX) { if (!QEMU_IS_ALIGNED(max_data, unit * base) || !QEMU_IS_ALIGNED(min_data, unit * base) { error_report("Invalid latency/bandwidth input."); exit(1); } base *= unit; }
[Prev in Thread] | Current Thread | [Next in Thread] |