/* * This file and its contents are supplied under the terms of the * Common Development and Distribution License ("CDDL"), version 1.0. * You may only use this file in accordance with the terms of version * 1.0 of the CDDL. * * A full copy of the text of the CDDL should have accompanied this * source. A copy of the CDDL is also available via the Internet at * http://www.illumos.org/license/CDDL. */ /* * Copyright 2023 Oxide Computer Company */ /* * AMD Zen Unified Memory Controller Driver * * This file forms the core logic around transforming a physical address that * we're used to using into a specific location on a DIMM. This has support for * a wide range of AMD CPUs and APUs ranging from Zen 1 - Zen 4. * * The goal of this driver is to implement the infrastructure and support * necessary to understand how DRAM requests are being routed in the system and * to be able to map those to particular channels and then DIMMs. This is used * as part of RAS (reliability, availability, and serviceability) to enable * aspects around understanding ECC errors, hardware topology, and more. Like * with any software project, there is more to do here. Please see the Future * Work section at the end of this big theory statement for more information. * * ------------------- * Driver Organization * ------------------- * * This driver is organized into two major pieces: * * 1. Logic to interface with hardware, discover the data fabric, memory * controller configuration, and transform that into a normalized fashion * that can be used across all different Zen family CPUs. This is * implemented generally in this file, and is designed to assume it is in * the kernel (as it requires access to the SMN, DF PCI registers, and the * amdzen nexus driver client services). * * 2. Logic that can take the above normalized memory information and perform * decoding (e.g. physical address to DIMM information). This generally * lives in common/mc/zen_uc/zen_umc_decode.c. This file is in common/, * meaning it is designed to be shared by userland and the kernel. Even * more so, it is designed to operate on a const version of our primary * data structure (zen_umc_t), not allowing it to be modified. This allows * us to more easily unit test the decoding logic and utilize it in other * circumstances such as with the mcdecode utility. * * There is corresponding traditional dev_ops(9S) and cb_ops(9S) logic in the * driver (currently this file) which take care of interfacing with the broader * operating system environment. * * There is only ever one instance of this driver, e.g. it is a singleton in * design pattern parlance. There is a single struct, the zen_umc_t found in the * global (albeit static) variable zen_umc. This structure itself contains a * hierarchical set of structures that describe the system. To make management * of memory simpler, all of the nested structures that we discover from * hardware are allocated in the same structure. The only exception to this rule * is when we cache serialized nvlists for dumping. * * The organization of the structures inside the zen_umc_t, generally mimics the * hardware organization and is structured as follows: * * +-----------+ * | zen_umc_t | * +-----------+ * | * +-------------------------------+ * v v * +--------------+ +--------------+ One instance of the * | zen_umc_df_t | ... | zen_umc_df_t | zen_umc_df_t per * +--------------+ +--------------+ discovered DF. * ||| * ||| * ||| +----------------+ +----------------+ Global DRAM * ||+--->| df_dram_rule_t | ... | df_dram_rule_t | rules for the * || +----------------+ +----------------+ platform. * || * || +--------------------+ +--------------------+ UMC remap * |+--->| zen_umc_cs_remap_t | ... | zen_umc_cs_remap_t | rule arrays. * | +--------------------+ +--------------------+ * | * v * +----------------+ +----------------+ One structure per * | zen_umc_chan_t | ... | zen_umc_chan_t | discovered DDR4/5 * +----------------+ +----------------+ memory channel. * |||| * |||| * |||| +----------------+ +----------------+ Channel specific * |||+--->| df_dram_rule_t | ... | df_dram_rule_t | copy of DRAM rules. * ||| +----------------+ +----------------+ Less than global. * ||| * ||| +---------------+ +---------------+ Per-Channel DRAM * ||+---->| chan_offset_t | ... | chan_offset_t | offset that is used * || +---------------+ +---------------+ for normalization. * || * || +-----------------+ Channel-specific * |+----->| umc_chan_hash_t | hashing rules. * | +-----------------+ * | * | +------------+ +------------+ One structure for * +------>| umc_dimm_t | ... | umc_dimm_t | each DIMM in the * +------------+ +------------+ channel. Always two. * | * | +----------+ +----------+ Per chip-select * +---> | umc_cs_t | ... | umc_cs_t | data. Always two. * +----------+ +----------+ * * In the data structures themselves you'll often find several pieces of data * that have the term 'raw' in their name. The point of these is to basically * capture the original value that we read from the register before processing * it. These are generally used either for debugging or to help answer future * curiosity with resorting to the udf and usmn tooling, which hopefully aren't * actually installed on systems. * * With the exception of some of the members in the zen_umc_t that are around * management of state for userland ioctls, everything in the structure is * basically write-once and from that point on should be treated as read-only. * * --------------- * Memory Decoding * --------------- * * To understand the process of memory decoding, it's worth going through and * understanding a bunch of the terminology that is used in this process. As an * additional reference when understanding this, you may want to turn to either * an older generation AMD BIOS and Kernel Developer's Guide or the more current * Processor Programming Reference. In addition, the imc driver, which is the * Intel equivalent, also provides an additional bit of reference. * * SYSTEM ADDRESS * * This is a physical address and is the way that the operating system * normally thinks of memory. System addresses can refer to many different * things. For example, you have traditional DRAM, memory-mapped PCIe * devices, peripherals that the processor exposes such as the xAPIC, data * from the FCH (Fusion Controller Hub), etc. * * TOM, TOM2, and the DRAM HOLE * * Physical memory has a complicated layout on x86 in part because of * support for traditional 16-bit and 32-bit systems. As a result, contrary * to popular belief, DRAM is not at a consistent address range in the * processor. AMD processors have a few different ranges. There is a 32-bit * region that starts at effectively physical address zero and goes to the * TOM MSR (top of memory -- Core::X86::Msr::TOP_MEM). This indicates a * limit below 4 GiB, generally around 2 GiB. * * From there, the next region of DRAM starts at 4 GiB and goes to TOM2 * (top of memory 2 -- Core::X86::Msr::TOM2). The region between TOM and * 4 GiB is called the DRAM hole. Physical addresses in this region are * used for memory mapped I/O. This breaks up contiguous physical * addresses being used for DRAM, creating a "hole". * * DATA FABRIC * * The data fabric (DF) is the primary interface that different parts of * the system use to communicate with one another. This includes the I/O * engines (where PCIe traffic goes), CPU caches and their cores, memory * channels, cross-socket communication, and a whole lot more. The first * part of decoding addresses and figuring out which DRAM channel an * address should be directed to all come from the data fabric. * * The data fabric is comprised of instances. So there is one instance for * each group of cores, each memory channel, etc. Each instance has its own * independent set of register information. As the data fabric is a series * of devices exposed over PCI, if you do a normal PCI configuration space * read or write that'll end up broadcasting the I/O. Instead, to access a * particular instance's register information there is an indirect access * mechanism. The primary way that this driver accesses data fabric * registers is via these indirect reads. * * There is one instance of the Data Fabric per socket starting with Zen 2. * In Zen 1, there was one instance of the data fabric per CCD -- core * complex die (see cpuid.c's big theory statement for more information). * * DF INSTANCE ID * * A DF instance ID is an identifier for a single entity or component in a * data fabric. The set of instance IDs is unique only with a single data * fabric. So for example, each memory channel, I/O endpoint (e.g. PCIe * logic), group of cores, has its own instance ID. Anything within the * same data fabric (e.g. the same die) can be reached via its instance ID. * The instance ID is used to indicate which instance to contact when * performing indirect accesses. * * Not everything that has an instance ID will be globally routable (e.g. * between multiple sockets). For things that are, such as the memory * channels and coherent core initiators, there is a second ID called a * fabric ID. * * DF FABRIC ID * * A DF fabric ID is an identifier that combines information to indicate * both which instance of the data fabric a component is on and a component * itself. So with this number you can distinguish between a memory channel * on one of two sockets. A Fabric ID is made up of two parts. The upper * part indicates which DF we are talking to and is referred to as a Node * ID. The Node ID is itself broken into two parts: one that identifies a * socket, and one that identifies a die. The lower part of a fabric ID is * called a component ID and indicates which component in a particular data * fabric that we are talking to. While only a subset of the total * components in the data fabric are routable, for everything that is, its * component ID matches its instance ID. * * Put differently, the component portion of a fabric ID and a component's * instance ID are always the same for routable entities. For things which * cannot be routed, they only have an instance ID and no fabric ID. * Because this code is always interacting with data fabric components that * are routable, sometimes instance ID and the component ID portion of the * data fabric ID may be used interchangeably. * * Finally, it's worth calling out that the number of bits that are used to * indicate the socket, die, and component in a fabric ID changes from * hardware generation to hardware generation. * * Inside the code here, the socket and die decomposition information is * always relative to the node ID. AMD phrases the decomposition * information in terms of a series of masks and shifts. This is * information that can be retrieved from the data fabric itself, allowing * us to avoid hardcoding too much information other than which registers * actually have which fields. With both masks and shifts, it's important * to establish which comes first. We follow AMD's convention and always * apply masks before shifts. With that, let's look at an example of a * made up bit set: * * Assumptions (to make this example simple): * o The fabric ID is 16 bits * o The component ID is 8 bits * o The node ID is 8 bits * o The socket and die ID are both 4 bits * * Here, let's say that we have the ID 0x2106. This decomposes into a * socket 0x2, die 0x1, and component 0x6. Here is how that works in more * detail: * * 0x21 0x06 * |------| |------| * Node ID Component ID * Mask: 0xff00 0x00ff * Shift: 8 0 * * Next we would decompose the Node ID as: * 0x2 0x1 * |------| |------| * Sock ID Die ID * Mask: 0xf0 0x0f * Shift: 4 0 * * Composing a fabric ID from its parts would work in a similar way by * applying masks and shifts. * * NORMAL ADDRESS * * A normal address is one of the primary address types that AMD uses in * memory decoding. It takes into account the DRAM hole, interleave * settings, and is basically the address that is dispatched to the broader * data fabric towards a particular DRAM channel. * * Often, phrases like 'normalizing the address' or normalization refer to * the process of transforming a system address into the channel address. * * INTERLEAVING * * The idea of interleaving is to take a contiguous range and weave it * between multiple different actual entities. Generally certain bits in * the range are used to select one of several smaller regions. For * example, if you have 8 regions each that are 4 GiB in size, that creates * a single 32 GiB region. You can use three bits in that 32 GiB space to * select one of the 8 regions. For a more visual example, see the * definition of this in uts/intel/io/imc/imc.c. * * CHANNEL * * A channel is used to refer to a single memory channel. This is sometimes * called a DRAM channel as well. A channel operates in a specific mode * based on the JEDEC DRAM standards (e.g. DDR4, LPDDR5, etc.). A * (LP)DDR4/5 channel may support up to two DIMMs inside the channel. The * number of slots is platform dependent and from there the number of DIMMs * installed can vary. Generally speaking, a DRAM channel defines a set * number of signals, most of which go to all DIMMs in the channel, what * varies is which "chip-select" is activated which causes a given DIMM to * pay attention or not. * * DIMM * * A DIMM refers to a physical hardware component that is installed into a * computer to provide access to dynamic memory. Originally this stood for * dual-inline memory module, though the DIMM itself has evolved beyond * that. A DIMM is organized into various pages, which are addressed by * a combination of rows, columns, banks, bank groups, and ranks. How this * fits together changes from generation to generation and is standardized * in something like DDR4, LPDDR4, DDR5, LPDDR5, etc. These standards * define the general individual modules that are assembled into a DIMM. * There are slightly different standards for combined memory modules * (which is what we use the term DIMM for). Examples of those include * things like registered DIMMs (RDIMMs). * * A DDR4 DIMM contains a single channel that is 64-bits wide with 8 check * bits. A DDR5 DIMM has a notable change in this scheme from earlier DDR * standards. It breaks a single DDR5 DIMM into two sub-channels. Each * sub-channel is independently addressed and contains 32-bits of data and * 8-bits of check data. * * ROW AND COLUMN * * The most basic building block of a DIMM is a die. A DIMM consists of * multiple dies that are organized together (we'll discuss the * organization next). A given die is organized into a series of rows and * columns. First, one selects a row. At which point one is able to select * a specific column. It is more expensive to change rows than columns, * leading a given row to contain approximately 1 KiB of data spread across * its columns. The exact size depends on the device. Each row/column is a * series of capacitors and transistors. The transistor is used to select * data from the capacitor and the capacitor actually contains the logical * 0/1 value. * * BANKS AND BANK GROUPS * * An individual DRAM die is organized in something called a bank. A DIMM * has a number of banks that sit in series. These are then grouped into * larger bank groups. Generally speaking, each bank group has the same * number of banks. Let's take a look at an example of a system with 4 * bank groups, each with 4 banks. * * +-----------------------+ +-----------------------+ * | Bank Group 0 | | Bank Group 1 | * | +--------+ +--------+ | | +--------+ +--------+ | * | | Bank 0 | | Bank 1 | | | | Bank 0 | | Bank 1 | | * | +--------+ +--------+ | | +--------+ +--------+ | * | +--------+ +--------+ | | +--------+ +--------+ | * | | Bank 2 | | Bank 3 | | | | Bank 2 | | Bank 3 | | * | +--------+ +--------+ | | +--------+ +--------+ | * +-----------------------+ +-----------------------+ * * +-----------------------+ +-----------------------+ * | Bank Group 2 | | Bank Group 3 | * | +--------+ +--------+ | | +--------+ +--------+ | * | | Bank 0 | | Bank 1 | | | | Bank 0 | | Bank 1 | | * | +--------+ +--------+ | | +--------+ +--------+ | * | +--------+ +--------+ | | +--------+ +--------+ | * | | Bank 2 | | Bank 3 | | | | Bank 2 | | Bank 3 | | * | +--------+ +--------+ | | +--------+ +--------+ | * +-----------------------+ +-----------------------+ * * On a DIMM, only a single bank and bank group can be active at a time for * reading or writing an 8 byte chunk of data. However, these are still * pretty important and useful because of the time involved to switch * between them. It is much cheaper to switch between bank groups than * between banks and that time can be cheaper than activating a new row. * This allows memory controllers to pipeline this substantially. * * RANK AND CHIP-SELECT * * The next level of organization is a rank. A rank is effectively an * independent copy of all the bank and bank groups on a DIMM. That is, * there are additional copies of the DIMM's organization, but not the data * itself. Originally a * single or dual rank DIMM was built such that one copy of everything was * on each physical side of the DIMM. As the number of ranks has increased * this has changed as well. Generally speaking, the contents of the rank * are equivalent. That is, you have the same number of bank groups, banks, * and each bank has the same number of rows and columns. * * Ranks are selected by what's called a chip-select, often abbreviated as * CS_L in the various DRAM standards. AMD also often abbreviates this as a * CS (which is not to be confused with the DF class of device called a * CS). These signals are used to select a rank to activate on a DIMM. * There are some number of these for each DIMM which is how the memory * controller chooses which of the DIMMs it's actually going to activate in * the system. * * One interesting gotcha here is how AMD organizes things. Each DIMM * logically is broken into two chip-selects in hardware. Between DIMMs * with more than 2 ranks and 3D stacked RDIMMs, there are ways to * potentially activate more bits. Ultimately these are mapped to a series * of rank multiplication logic internally. These ultimately then control * some of these extra pins, though the exact method isn't 100% clear at * this time. * * ----------------------- * Rough Hardware Process * ----------------------- * * To better understand how everything is implemented and structured, it's worth * briefly describing what happens when hardware wants to read a given physical * address. This is roughly summarized in the following chart. In the left hand * side is the type of address, which is transformed and generally shrinks along * the way. Next to it is the actor that is taking action and the type of * address that it starts with. * * +---------+ +------+ * | Virtual | | CPU | * | Address | | Core | * +---------+ +------+ * | | The CPU core receives a memory request and then * | * . . . . determines whether this request is DRAM or MMIO * | | (memory-mapped I/O) and then sends it to the data * v v fabric. * +----------+ +--------+ * | Physical | | Data | * | Address | | Fabric | * +----------+ +--------+ * | | The data fabric instance in the CCX/D uses the * | * . . . . programmed DRAM rules to determine what DRAM * | | channel to direct a request to and what the * | | channel-relative address is. It then sends the * | | request through the fabric. Note, the number of * | | DRAM rules varies based on the processor SoC. * | | Server parts like Milan have many more rules than * | | an APU like Cezanne. The DRAM rules tell us both * v v how to find and normalize the physical address. * +---------+ +---------+ * | Channel | | DRAM | * | Address | | Channel | * +---------+ +---------+ * | | The UMC (unified memory controller) receives the * | * . . . . DRAM request and determines which DIMM to send * | | the request to along with the rank, banks, row, * | | column, etc. It initiates a DRAM transaction and * | | then sends the results back through the data * v v fabric to the CPU core. * +---------+ +--------+ * | DIMM | | Target | * | Address | | DIMM | * +---------+ +--------+ * * The above is all generally done in hardware. There are multiple steps * internal to this that we end up mimicking in software. This includes things * like, applying hashing logic, address transformations, and related. * Thankfully the hardware is fairly generic and programmed with enough * information that we can pull out to figure this out. The rest of this theory * statement covers the major parts of this: interleaving, the act of * determining which memory channel to actually go to, and normalization, the * act of removing some portion of the physical address bits to determine the * address relative to a channel. * * ------------------------ * Data Fabric Interleaving * ------------------------ * * One of the major parts of address decoding is to understand how the * interleaving features work in the data fabric. This is used to allow an * address range to be spread out between multiple memory channels and then, * later on, when normalizing the address. As mentioned above, a system address * matches a rule which has information on interleaving. Interleaving comes in * many different flavors. It can be used to just switch between channels, * sockets, and dies. It can also end up involving some straightforward and some * fairly complex hashing operations. * * Each DRAM rule has instructions on how to perform this interleaving. The way * this works is that the rule first says to start at a given address bit, * generally ranging from bit 8-12. These influence the granularity of the * interleaving going on. From there, the rules determine how many bits to use * from the address to determine the die, socket, and channel. In the simplest * form, these perform a log2 of the actual number of things you're interleaving * across (we'll come back to non-powers of two). So let's work a few common * examples: * * o 8-channel interleave, 1-die interleave, 2-socket interleave * Start at bit 9 * * In this case we have 3 bits that determine the channel to use, 0 bits * for the die, 1 bit for the socket. Here we would then use the following * bits to determine what the channel, die, and socket IDs are: * * [12] - Socket ID * [11:9] - Channel ID * * You'll note that there was no die-interleave, which means the die ID is * always zero. This is the general thing you expect to see in Zen 2 and 3 * based systems as they only have one die or a Zen 1 APU. * * o 2-channel interleave, 4-die interleave, 2-socket interleave * Start at bit 10 * * In this case we have 1 bit for the channel and socket interleave. We * have 2 bits for the die. This is something you might see on a Zen 1 * system. This results in the following bits: * * [13] - Socket ID * [12:11] - Die ID * [10] - Channel ID * * * COD and NPS HASHING * * However, this isn't the only primary extraction rule of the above values. The * other primary method is using a hash. While the exact hash methods vary * between Zen 2/3 and Zen 4 based systems, they follow a general scheme. In the * system there are three interleaving configurations that are either global or * enabled on a per-rule basis. These indicate whether one should perform the * XOR computation using addresses at: * * o 64 KiB (starting at bit 16) * o 2 MiB (starting at bit 21) * o 1 GiB (starting at bit 30) * * In this world, you take the starting address bit defined by the rule and XOR * it with each enabled interleave address. If you have more than one bit to * select (e.g. because you are hashing across more than 2 channels), then you * continue taking subsequent bits from each enabled region. So the second bit * would use 17, 21, and 31 if all three ranges were enabled while the third bit * would use 18, 22, and 32. While these are straightforward, there is a catch. * * While the DRAM rule contains what the starting address bit, you don't * actually use subsequent bits in the same way. Instead subsequent bits are * deterministic and use bits 12 and 13 from the address. This is not the same * consecutive thing that one might expect. Let's look at a Rome/Milan based * example: * * o 8-channel "COD" hashing, starting at address 9. All three ranges enabled. * 1-die and 1-socket interleaving. * * In this model we are using 3 bits for the channel, 0 bits for the socket * and die. * * Channel ID[0] = addr[9] ^ addr[16] ^ addr[21] ^ addr[30] * Channel ID[1] = addr[12] ^ addr[17] ^ addr[22] ^ addr[31] * Channel ID[2] = addr[13] ^ addr[18] ^ addr[23] ^ addr[32] * * So through this scheme we'd have a socket/die of 0, and then the channel * ID is computed based on that. The number of bits that we use here * depends on how many channels the hash is going across. * * The Genoa and related variants, termed "NPS", has a few wrinkles. First, * rather than 3 bits being used for the channel, up to 4 bits are. Second, * while the Rome/Milan "COD" hash above does not support socket or die * interleaving, the "NPS" hash actually supports socket interleaving. However, * unlike the straightforward non-hashing scheme, the first bit is used to * determine the socket when enabled as opposed to the last one. In addition, if * we're not performing socket interleaving, then we end up throwing address bit * 14 into the mix here. Let's look at examples: * * o 4-channel "NPS" hashing, starting at address 8. All three ranges enabled. * 1-die and 1-socket interleaving. * * In this model we are using 2 bits for the channel, 0 bits for the socket * and die. Because socket interleaving is not being used, bit 14 ends up * being added into the first bit of the channel selection. Presumably this * is to improve the address distribution in some form. * * Channel ID[0] = addr[8] ^ addr[16] ^ addr[21] ^ addr[30] ^ addr[14] * Channel ID[1] = addr[12] ^ addr[17] ^ addr[22] ^ addr[31] * * o 8-channel "NPS" hashing, starting at address 9. All three ranges enabled. * 1-die and 2-socket interleaving. * * In this model we are using 3 bits for the channel and 1 for the socket. * The die is always set to 0. Unlike the above, address bit 14 is not used * because it ends up being required for the 4th address bit. * * Socket ID[0] = addr[9] ^ addr[16] ^ addr[21] ^ addr[30] * Channel ID[0] = addr[12] ^ addr[17] ^ addr[22] ^ addr[31] * Channel ID[1] = addr[13] ^ addr[18] ^ addr[23] ^ addr[32] * Channel ID[2] = addr[14] ^ addr[19] ^ addr[24] ^ addr[33] * * * ZEN 3 6-CHANNEL * * These were the simple cases. Things get more complex when we move to * non-power of 2 based hashes between channels. There are two different sets of * these schemes. The first of these is 6-channel hashing that was added in Zen * 3. The second of these is a more complex and general form that was added in * Zen 4. Let's start with the Zen 3 case. The Zen 3 6-channel hash requires * starting at address bits 11 or 12 and varies its logic somewhat from there. * In the 6-channel world, the socket and die interleaving must be disabled. * Let's walk through an example: * * o 6-channel Zen 3, starting at address 11. 2M and 1G range enabled. * 1-die and 1-socket interleaving. * * Regardless of the starting address, we will always use three bits to * determine a channel address. However, it's worth calling out that the * 64K range is not considered for this at all. Another oddity is that when * calculating the hash bits the order of the extracted 2M and 1G addresses * are different. * * This flow starts by calculating the three hash bits. This is defined * below. In the following, all bits marked with an '@' are ones that will * change when starting at address bit 12. In those cases the value will * increase by 1. Here's how we calculate the hash bits: * * hash[0] = addr[11@] ^ addr[14@] ^ addr[23] ^ addr[32] * hash[1] = addr[12@] ^ addr[21] ^ addr[30] * hash[2] = addr[13@] ^ addr[22] ^ addr[31] * * With this calculated, we always assign the first bit of the channel * based on the hash. The other bits are more complicated as we have to * deal with that gnarly power of two problem. We determine whether or not * to use the hash bits directly in the channel based on their value. If * they are not equal to 3, then we use it, otherwise if they are, then we * need to go back to the physical address and we take its modulus. * Basically: * * Channel Id[0] = hash[0] * if (hash[2:1] == 3) * Channel ID[2:1] = (addr >> [11@+3]) % 3 * else * Channel ID[2:1] = hash[2:1] * * * ZEN 4 NON-POWER OF 2 * * I hope you like modulus calculations, because things get even more complex * here now in Zen 4 which has many more modulus variations. These function in a * similar way to the older 6-channel hash in Milan. They require one to start * at address bit 8, they require that there is no die interleaving, and they * support socket interleaving. The different channel arrangements end up in one * of two sets of modulus values: a mod % 3 and a mod % 5 based on the number * of channels used. Unlike the Milan form, all three address ranges (64 KiB, 2 * MiB, 1 GiB) are allowed to be used. * * o 6-channel Zen 4, starting at address 8. 64K, 2M, and 1G range enabled. * 1-die and 2-socket interleaving. * * We start by calculating the following set of hash bits regardless of * the number of channels that exist. The set of hash bits that is actually * used in various computations ends up varying based upon the number of * channels used. In 3-5 configs, only hash[0] is used. 6-10, both hash[0] * and hash[2] (yes, not hash[1]). The 12 channel config uses all three. * * hash[0] = addr[8] ^ addr[16] ^ addr[21] ^ addr[30] ^ addr[14] * hash[1] = addr[12] ^ addr[17] ^ addr[22] ^ addr[31] * hash[2] = addr[13] ^ addr[18] ^ addr[23] ^ addr[32] * * Unlike other schemes where bits directly map here, they instead are used * to seed the overall value. Depending on whether hash[0] is a 0 or 1, the * system goes through two different calculations entirely. Though all of * them end up involving the remainder of the system address going through * the modulus. In the following, a '3@' indicates the modulus value would * be swapped to 5 in a different scenario. * * Channel ID = addr[63:14] % 3@ * if (hash[0] == 1) * Channel ID = (Channel ID + 1) % 3@ * * Once this base has for the channel ID has been calculated, additional * portions are added in. As this is the 6-channel form, we say: * * Channel ID = Channel ID + (hash[2] * 3@) * * Finally the socket is deterministic and always comes from hash[0]. * Basically: * * Socket ID = hash[0] * * o 12-channel Zen 4, starting at address 8. 64K, 2M, and 1G range enabled. * 1-die and 1-socket interleaving. * * This is a variant of the above. The hash is calculated the same way. * The base Channel ID is the same and if socket interleaving were enabled * it would also be hash[0]. What instead differs is how we use hash[1] * and hash[2]. The following logic is used instead of the final * calculation above. * * Channel ID = Channel ID + (hash[2:1] * 3@) * * * POST BIT EXTRACTION * * Now, all of this was done to concoct up a series of indexes used. However, * you'll note that a given DRAM rule actually already has a fabric target. So * what do we do here? We add them together. * * The data fabric has registers that describe which bits in a fabric ID * correspond to a socket, die, and channel. Taking the channel, die, and socket * IDs above, one can construct a fabric ID. From there, we add the two data * fabric IDs together and can then get to the fabric ID of the actual logical * target. This is why all of the socket and die interleaving examples with no * interleaving are OK to result in a zero. The idea here is that the base * fabric ID in the DRAM rule will take care of indicating those other things as * required. * * You'll note the use of the term "logical target" up above. That's because * some platforms have the ability to remap logical targets to physical targets * (identified by the use of the ZEN_UMC_FAM_F_TARG_REMAP flag in the family * data). The way that remapping works changes based on the hardware generation. * This was first added in Milan (Zen 3) CPUs. In that model, you would use the * socket and component information from the target ID to identify which * remapping rules to use. On Genoa (Zen 4) CPUs, you would instead use * information in the rule itself to determine which of the remap rule sets to * use and then uses the component ID to select which rewrite rule to use. * * Finally, there's one small wrinkle with this whole scheme that we haven't * discussed: what actually is the address that we plug into this calculation. * While you might think it actually is just the system address itself, that * isn't actually always the case. Sometimes rather than using the address * itself, it gets normalized based on the DRAM rule, which involves subtracting * out the base address and potentially subtracting out the size of the DRAM * hole (if the address is above the hole and hoisting is active for that * range). When this is performed appears to tie to the DF generation. After Zen * 3, it is always the default (e.g. Zen 4 and things from DF gen 3.5). At and * before Zen 3, it only occurs if we are doing a non-power of 2 based hashing. * * -------------------------------------------- * Data Fabric Interleave Address Normalization * -------------------------------------------- * * While you may have thought that we were actually done with the normalization * fun in the last section, there's still a bit more here that we need to * consider. In particular, there's a secondary transformation beyond * interleaving that occurs as part of constructing the channel normalized * address. Effectively, we need to account for all the bits that were used in * the interleaving and generally speaking remove them from our normalized * address. * * While this may sound weird on paper, the way to think about it is that * interleaving at some granularity means that each device is grabbing the same * set of addresses, the interleave just is used to direct it to its own * location. When working with a channel normalized address, we're effectively * creating a new region of addresses that have meaning within the DIMMs * themselves. The channel doesn't care about what got it there, mainly just * what it is now. So with that in mind, we need to discuss how we remove all * the interleaving information in our different modes. * * Just to make sure it's clear, we are _removing_ all bits that were used for * interleaving. This causes all bits above the removed ones to be shifted * right. * * First, we have the case of standard power of 2 interleaving that applies to * the 1, 2, 4, 8, 16, and 32 channel configurations. Here, we need to account * for the total number of bits that are used for the channel, die, and socket * interleaving and we simply remove all those bits starting from the starting * address. * * o 8-channel interleave, 1-die interleave, 2-socket interleave * Start at bit 9 * * If we look at this example, we are using 3 bits for the channel, 1 for * the socket, for a total of 4 bits. Because this is starting at bit 9, * this means that interleaving covers the bit range [12:9]. In this case * our new address would be (orig[63:13] >> 4) | orig[8:0]. * * * COD and NPS HASHING * * That was the simple case, next we have the COD/NPS hashing case that we need * to consider. If we look at these, the way that they work is that they split * which bits they use for determining the channel address and then hash others * in. Here, we need to extract the starting address bit, then continue at bit * 12 based on the number of bits in use and whether or not socket interleaving * is at play for the NPS variant. Let's look at an example here: * * o 8-channel "COD" hashing, starting at address 9. All three ranges enabled. * 1-die and 1-socket interleaving. * * Here we have three total bits being used. Because we start at bit 9, this * means we need to drop bits [13:12], [9]. So our new address would be: * * orig[63:14] >> 3 | orig[11:10] >> 1 | orig[8:0] * | | +-> stays the same * | +-> relocated to bit 9 -- shifted by 1 because we * | removed bit 9. * +--> Relocated to bit 11 -- shifted by 3 because we removed bits, 9, 12, * and 13. * * o 8-channel "NPS" hashing, starting at address 8. All three ranges enabled. * 1-die and 2-socket interleaving. * * Here we need to remove bits [14:12], [8]. We're removing an extra bit * because we have 2-socket interleaving. This results in a new address of: * * orig[63:15] >> 4 | orig[11:9] >> 1 | orig[7:0] * | | +-> stays the same * | +-> relocated to bit 8 -- shifted by 1 because we * | removed bit 8. * +--> Relocated to bit 11 -- shifted by 4 because we removed bits, 8, 12, * 13, and 14. * * * ZEN 3 6-CHANNEL * * Now, to the real fun stuff, our non-powers of two. First, let's start with * our friend, the Zen 3 6-channel hash. So, the first thing that we need to do * here is start by recomputing our hash again based on the current normalized * address. Regardless of the hash value, this first removes all three bits from * the starting address, so that's removing either [14:12] or [13:11]. * * The rest of the normalization process here is quite complex and somewhat mind * bending. Let's start working through an example here and build this up. * First, let's assume that each channel has a single 16 GiB RDIMM. This would * mean that the channel itself has 96 GiB RDIMM. However, by removing 3 bits * worth, that technically corresponds to an 8-channel configuration that * normally suggest a 128 GiB configuration. The processor requires us to record * this fact in the DF::Np2ChannelConfig register. The value that it wants us a * bit weird. We believe it's calculated by the following: * * 1. Round the channel size up to the next power of 2. * 2. Divide this total size by 64 KiB. * 3. Determine the log base 2 that satisfies this value. * * In our particular example above. We have a 96 GiB channel, so for (1) we end * up with 128 GiB (2^37). We now divide that by 64 KiB (2^16), so this becomes * 2^(37 - 16) or 2^21. Because we want the log base 2 of 2^21 from (2), this * simply becomes 21. The DF::Np2ChannelConfig has two members, a 'space 0' and * 'space 1'. Near as we can tell, in this mode only 'space 0' is used. * * Before we get into the actual normalization scheme, we have to ask ourselves * how do we actually interleave data 6 ways. The scheme here is involved. * First, it's important to remember like with other normalization schemes, we * do adjust for the address for the base address in the DRAM rule and then also * take into account the DRAM hole if present. * * If we delete 3 bits, let's take a sample address and see where it would end * up in the above scheme. We're going to take our 3 address bits and say that * they start at bit 12, so this means that the bits removed are [14:12]. So the * following are the 8 addresses that we have here and where they end up * starting with 1ff: * * o 0x01ff -> 0x1ff, Channel 0 (hash 0b000) * o 0x11ff -> 0x1ff, Channel 1 (hash 0b001) * o 0x21ff -> 0x1ff, Channel 2 (hash 0b010) * o 0x31ff -> 0x1ff, Channel 3 (hash 0b011) * o 0x41ff -> 0x1ff, Channel 4 (hash 0b100) * o 0x51ff -> 0x1ff, Channel 5 (hash 0b101) * o 0x61ff -> 0x3000001ff, Channel 0 (hash 0b110) * o 0x71ff -> 0x3000001ff, Channel 1 (hash 0b111) * * Yes, we did just jump to near the top of what is a 16 GiB DIMM's range for * those last two. The way we determine when to do this jump is based on our * hash. Effectively we ask what is hash[2:1]. If it is 0b11, then we need to * do something different and enter this special case, basically jumping to the * top of the range. If we think about a 6-channel configuration for a moment, * the thing that doesn't exist are the traditional 8-channel hash DIMMs 0b110 * and 0b111. * * If you go back to the interleave this kind of meshes, that tried to handle * the case of the hash being 0, 1, and 2, normally, and then did special things * with the case of the hash being in this upper quadrant. The hash then * determined where it went by shifting over the upper address and doing a mod * 3 and using that to determine the upper two bits. With that weird address at * the top of the range, let's go through and see what else actually goes to * those weird addresses: * * o 0x08000061ff -> 0x3000001ff, Channel 2 (hash 0b110) * o 0x08000071ff -> 0x3000001ff, Channel 3 (hash 0b111) * o 0x10000061ff -> 0x3000001ff, Channel 4 (hash 0b110) * o 0x10000071ff -> 0x3000001ff, Channel 5 (hash 0b111) * * Based on the above you can see that we've split the 16 GiB DIMM into a 12 GiB * region (e.g. [ 0x0, 0x300000000 ), and a 4 GiB region [ 0x300000000, * 0x400000000 ). What seems to happen is that the CPU algorithmically is going * to put things in this upper range. To perform that action it goes back to the * register information that we stored in DF::Np2ChannelConfig. The way this * seems to be thought of is it wants to set the upper two bits of a 64 KiB * chunk (e.g. bits [15:14]) to 0b11 and then shift that over based on the DIMM * size. * * Our 16 GiB DIMM has 34 bits, so effectively we want to set bits [33:32] in * this case. The channel is 37 bits wide, which the CPU again knows as 2^21 * * 2^16. So it constructs the 64 KiB value of [15:14] = 0b11 and fills the rest * with zeros. It then multiplies it by 2^(21 - 3), or 2^18. The - 3 comes from * the fact that we removed 3 address bits. This when added to the above gets * us bits [33,32] = 0b11. * * While this appears to be the logic, I don't have a proof that this scheme * actually evenly covers the entire range, but a few examples appear to work * out. * * With this, the standard example flow that we give, results in something like: * * o 6-channel Zen 3, starting at address 11. 2M and 1G range enabled. Here, * we assume that the value of the NP2 space0 is 21 bits. This example * assumes we have 96 GiB total memory, which means rounding up to 128 GiB. * * Step 1 here is to adjust our address to remove the three bits indicated. * So we simply always set our new address to: * * orig[63:14] >> 3 | orig[10:0] * | +-> stays the same * +--> Relocated to bit 11 because a 6-channel config always uses 3 bits to * perform interleaving. * * At this step, one would need to consult the hash of the normalized * address before removing bits (but after adjusting for the base / DRAM * hole). If hash[2:1] == 3, then we would say that the address is actually: * * 0b11 << 32 | orig[63:14] >> 3 | orig[10:0] * * * ZEN 4 NON-POWER OF 2 * * Next, we have the DFv4 versions of the 3, 5, 6, 10, and 12 channel hashing. * An important part of this is whether or not there is any socket hashing going * on. Recall there, that if socket hashing was going on, then it is part of the * interleave logic; however, if it is not, then its hash actually becomes * part of the normalized address, but not in the same spot! * * In this mode, we always remove the bits that are actually used by the hash. * Recall that some modes use hash[0], others hash[0] and hash[2], and then only * the 12-channel config uses hash[2:0]. This means we need to be careful in how * we actually remove address bits. All other bits in this lower range we end up * keeping and using. The top bits, e.g. addr[63:14] are kept and divided by the * actual channel-modulus. If we're not performing socket interleaving and * therefore need to keep the value of hash[0], then it is appended as the least * significant bit of that calculation. * * Let's look at an example of this to try to make sense of it all. * * o 6-channel Zen 4, starting at address 8. 64K, 2M, and 1G range enabled. * 1-die and 2-socket interleaving. * * Here we'd start by calculating hash[2:0] as described in the earlier * interleaving situation. Because we're using a socket interleave, we will * not opt to include hash[0] in the higher-level address calculation. * Because this is a 6-channel calculation, our modulus is 3. Here, we will * strip out bits 8 and 13 (recall in the interleaving 6-channel example we * ignored hash[1], thus no bit 12 here). Our new address will be: * * (orig[63:14] / 3) >> 2 | orig[12:9] >> 1 | orig[7:0] * | | +-> stays the same * | +-> relocated to bit 8 -- shifted by 1 because * | we removed bit 8. * +--> Relocated to bit 12 -- shifted by 2 because we removed bits 8 and * 13. * * o 12-channel Zen 4, starting at address 8. 64K, 2M, and 1G range enabled. * 1-die and 1-socket interleaving. * * This is a slightly different case from the above in two ways. First, we * will end up removing bits 8, 12, and 13, but then we'll also reuse * hash[0]. Our new address will be: * * ((orig[63:14] / 3) << 1 | hash[0]) >> 3 | orig[11:9] >> 1 | orig[7:0] * | | +-> stays the * | | same * | +-> relocated to bit 8 -- shifted by * | 1 because we removed bit 8. * +--> Relocated to bit 11 -- shifted by 3 because we removed bits 8, 12, * and 13. * * That's most of the normalization process for the time being. We will have to * revisit this when we have to transform a normal address into a system address * and undo all this. * * ------------------------------------- * Selecting a DIMM and UMC Organization * ------------------------------------- * * One of the more nuanced things in decoding and encoding is the question of * where do we send a channel normalized address. That is, now that we've gotten * to a given channel, we need to transform the address into something * meaningful for a DIMM, and select a DIMM as well. The UMC SMN space contains * a number of Base Address and Mask registers which they describe as activating * a chip-select. A given UMC has up to four primary chip-selects (we'll come * back to DDR5 sub-channels later). The first two always go to the first DIMM * in the channel and the latter two always go to the second DIMM in the * channel. Put another way, you can always determine which DIMM you are * referring to by taking the chip-select and shifting it by 1. * * The UMC Channel registers are organized a bit differently in different * hardware generations. In a DDR5 based UMC, almost all of our settings are on * a per-chip-select basis while as in a DDR4 based system only the bases and * masks are. While gathering data we normalize this such that each logical * chip-select (umc_cs_t) that we have in the system has the same data so that * way DDR4 and DDR5 based systems are the same to the decoding logic. There is * also channel-wide data such as hash configurations and related. * * Each channel has a set of base and mask registers (and secondary ones as * well). To determine if we activate a given one, we first check if the * enabled bit is set. The enabled bit is set on a per-base basis, so both the * primary and secondary registers have separate enables. As there are four of * each base, mask, secondary base, and secondary mask, we say that if a * normalized address matches either a given indexes primary or secondary index, * then it activates that given UMC index. The basic formula for an enabled * selection is: * * NormAddr & ~Mask[i] == Base[i] & ~Mask[i] * * Once this is selected, this index in the UMC is what it always used to derive * the rest of the information that is specific to a given chip-select or DIMM. * An important thing to remember is that from this point onwards, while there * is a bunch of hashing and interleaving logic it doesn't change which UMC * channel we read the data from. Though the particular DIMM, rank, and address * we access will change as we go through hashing and interleaving. * * ------------------------ * Row and Column Selection * ------------------------ * * The number of bits that are used for the row and column address of a DIMM * varies based on the type of module itself. These depend on the density of a * DIMM module, e.g. how large an individual DRAM block is, a value such as 16 * Gbit, and the number of these wide it is, which is generally phrased as X4, * X8, and X16. The memory controller encodes the number of bits (derived from * the DIMM's SPD data) and then determines which bits are used for addresses. * * Based on this information we can initially construct a row and a column * address by leveraging the information about the number of bits and then * extracting the correct bits out of the normalized channel address. * * If you've made it this far, you know nothing is quite this simple, despite it * seeming so. Importantly, not all DIMMs actually have storage that is a power * of 2. As such, there's another bit that we have to consult to transform the * actual value that we have for a row, remarkably the column somehow has no * transformations applied to it. * * The hardware gives us information on inverting the two 'most significant * bits' of the row address which we store in 'ucs_inv_msbs'. First, we have the * question of what are our most significant bits here. This is basically * determined by the number of low and high row bits. In this case higher * actually is what we want. Note, the high row bits only exist in DDR4. Next, * we need to know whether we used the primary or secondary base/mask pair for * this as there is a primary and secondary inversion bits. The higher bit of * the inversion register (e.g ucs_inv_msbs[1]) corresponds to the highest row * bit. A zero in the bit position indicates that we should not perform an * inversion where as a one says that we should invert this. * * To actually make this happen we can take advantage of the fact that the * meaning of a 0/1 above means that this can be implemented with a binary * exclusive-OR (XOR). Logically speaking if we have a don't invert setting * present, a 0, then x ^ 0 is always x. However, if we have a 1 present, then * we know that (for a single bit) x ^ 1 = ~x. We take advantage of this fact in * the row logic. * * --------------------- * Banks and Bank Groups * --------------------- * * While addressing within a given module is done by the use of a row and column * address, to increase storage density a module generally has a number of * banks, which may be organized into one or more bank groups. While a given * DDR4/5 access happens in some prefetched chunk of say 64 bytes (what do you * know, that's a cacheline), that all occurs within a single bank. The addition * of bank groups makes it easier to access data in parallel -- it is often * faster to read from another bank group than to read another region inside a * bank group. * * Based on the DIMMs internal configuration, there will be a specified number * of bits used for the overall bank address (including bank group bits) * followed by a number of bits actually used for bank groups. There are * separately an array of bits used to concoct the actual address. It appears, * mostly through experimental evidence, that the bank group bits occur first * and then are followed by the bank selection itself. This makes some sense if * you assume that switching bank groups is faster than switching banks. * * So if we see the UMC noting 4 bank bits and 2 bank groups bits, that means * that the umc_cs_t's ucs_bank_bits[1:0] correspond to bank_group[1:0] and * ucs_bank_bits[3:2] correspond to bank_address[1:0]. However, if there were no * bank bits indicated, then all of the address bits would correspond to the * bank address. * * Now, this would all be straightforward if not for hashing, our favorite. * There are five bank hashing registers per channel (UMC_BANK_HASH_DDR4, * UMC_BANK_HASH_DDR5), one that corresponds to the five possible bank bits. To * do this we need to use the calculated row and column that we previously * determined. This calculation happens in a few steps: * * 1) First check if the enable bit is set in the rule. If not, just use the * normal bank address bit and we're done. * 2) Take a bitwise-AND of the calculated row and hash register's row value. * Next do the same thing for the column. * 3) For each bit in the row, progressively XOR it, e.g. row[0] ^ row[1] ^ * row[2] ^ ... to calculate a net bit value for the row. This then * repeats itself for the column. What basically has happened is that we're * using the hash register to select which bits to impact our decision. * Think of this as a traditional bitwise functional reduce. * 4) XOR the combined rank bit with the column bit and the actual bank * address bit from the normalized address. So if this were bank bit 0, * which indicated we should use bit 15 for bank[0], then we would * ultimately say our new bit is norm_addr[15] ^ row_xor ^ col_xor * * An important caveat is that we would only consult all this if we actually * were told that the bank bit was being used. For example if we had 3 bank * bits, then we'd only check the first 3 hash registers. The latter two would * be ignored. * * Once this process is done, then we can go back and split the activated bank * into the actual bank used and the bank group used based on the first bits * going to the bank group. * * --------------- * DDR5 Sub-channel * --------------- * * As described in the definitions section, DDR5 has the notion of a * sub-channel. Here, a single bit is used to determine which of the * sub-channels to actually operate and utilize. Importantly the same * chip-select seems to apply to both halves of a given sub-channel. * * There is also a hash that is used here. The hash here utilizes the calculated * bank, column, and row and follows the same pattern used in the bank * calculation where we do a bunch of running exclusive-ORs and then do that * with the original value we found to get the new value. Because there's only * one bit for the sub-channel, we only have a single hash to consider. * * ------------------------------------------- * Ranks, Chip-Select, and Rank Multiplication * ------------------------------------------- * * The notion of ranks and the chip-select are interwoven. From a strict DDR4 * RDIMM perspective, there are two lines that are dedicated for chip-selects * and then another two that are shared with three 'chip-id' bits that are used * in 3DS RDIMMs. In all cases the controller starts with two logical chip * selects and then uses something called rank multiplication to figure out how * to multiplex that and map to the broader set of things. Basically, in * reality, DDR4 RDIMMs allow for 4 bits to determine a rank and then 3DS RDIMMs * use 2 bits for a rank and 3 bits to select a stacked chip. In DDR5 this is * different and you just have 2 bits for a rank. * * It's not entirely clear from what we know from AMD, but it seems that we use * the RM bits as a way to basically go beyond the basic 2 bits of chip-select * which is determined based on which channel we logically activate. Initially * we treat this as two distinct things, here as that's what we get from the * hardware. There are two hashes here a chip-select and rank-multiplication * hash. Unlike the others, which rely on the bank, row, and column addresses, * this hash relies on the normalized address. So we calculate that mask and do * our same xor dance. * * There is one hash for each rank multiplication bit and chip-select bit. The * number of rank multiplication bits is given to us. The number of chip-select * bits is fixed, it's simply two because there are four base/mask registers and * logical chip-selects in a given UMC channel. The chip-select on some DDR5 * platforms has a secondary exclusive-OR hash that can be applied. As this only * exists in some families, for any where it does exist, we seed it to be zero * so that it becomes a no-op. * * ----------- * Future Work * ----------- * * As the road goes ever on and on, down from the door where it began, there are * still some stops on the journey for this driver. In particular, here are the * major open areas that could be implemented to extend what this can do: * * o The ability to transform a normalized channel address back to a system * address. This is required for MCA/MCA-X error handling as those generally * work in terms of channel addresses. * o Integrating with the MCA/MCA-X error handling paths so that way we can * take correct action in the face of ECC errors and allowing recovery from * uncorrectable errors. * o Providing memory controller information to FMA so that way it can opt to * do predictive failure or give us more information about what is fault * with ECC errors. * o Figuring out if we will get MCEs for privilged address decoding and if so * mapping those back to system addresses and related. * o 3DS RDIMMs likely will need a little bit of work to ensure we're handling * the resulting combination of the RM bits and CS and reporting it * intelligently. */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include static zen_umc_t *zen_umc; /* * Per-CPU family information that describes the set of capabilities that they * implement. When adding support for new CPU generations, you must go through * what documentation you have and validate these. The best bet is to find a * similar processor and see what has changed. Unfortunately, there really isn't * a substitute for just basically checking every register. The family name * comes from the amdzen_c_family(). One additional note for new CPUs, if our * parent amdzen nexus driver does not attach (because the DF has changed PCI * IDs or more), then just adding something here will not be sufficient to make * it work. */ static const zen_umc_fam_data_t zen_umc_fam_data[] = { { .zufd_family = X86_PF_AMD_NAPLES, .zufd_dram_nrules = 16, .zufd_cs_nrules = 2, .zufd_umc_style = ZEN_UMC_UMC_S_DDR4, .zufd_chan_hash = UMC_CHAN_HASH_F_BANK | UMC_CHAN_HASH_F_CS }, { .zufd_family = X86_PF_HYGON_DHYANA, .zufd_dram_nrules = 16, .zufd_cs_nrules = 2, .zufd_umc_style = ZEN_UMC_UMC_S_DDR4, .zufd_chan_hash = UMC_CHAN_HASH_F_BANK | UMC_CHAN_HASH_F_CS }, { .zufd_family = X86_PF_AMD_DALI, .zufd_dram_nrules = 2, .zufd_cs_nrules = 2, .zufd_umc_style = ZEN_UMC_UMC_S_DDR4_APU, .zufd_chan_hash = UMC_CHAN_HASH_F_BANK | UMC_CHAN_HASH_F_CS }, { .zufd_family = X86_PF_AMD_ROME, .zufd_flags = ZEN_UMC_FAM_F_NP2 | ZEN_UMC_FAM_F_NORM_HASH | ZEN_UMC_FAM_F_UMC_HASH, .zufd_dram_nrules = 16, .zufd_cs_nrules = 2, .zufd_umc_style = ZEN_UMC_UMC_S_DDR4, .zufd_chan_hash = UMC_CHAN_HASH_F_BANK | UMC_CHAN_HASH_F_RM | UMC_CHAN_HASH_F_CS }, { .zufd_family = X86_PF_AMD_RENOIR, .zufd_flags = ZEN_UMC_FAM_F_NORM_HASH, .zufd_dram_nrules = 2, .zufd_cs_nrules = 2, .zufd_umc_style = ZEN_UMC_UMC_S_DDR4_APU, .zufd_chan_hash = UMC_CHAN_HASH_F_BANK | UMC_CHAN_HASH_F_PC | UMC_CHAN_HASH_F_CS }, { .zufd_family = X86_PF_AMD_MATISSE, .zufd_flags = ZEN_UMC_FAM_F_NORM_HASH | ZEN_UMC_FAM_F_UMC_HASH, .zufd_dram_nrules = 16, .zufd_cs_nrules = 2, .zufd_umc_style = ZEN_UMC_UMC_S_DDR4, .zufd_chan_hash = UMC_CHAN_HASH_F_BANK | UMC_CHAN_HASH_F_RM | UMC_CHAN_HASH_F_CS }, { .zufd_family = X86_PF_AMD_VAN_GOGH, .zufd_flags = ZEN_UMC_FAM_F_NORM_HASH, .zufd_dram_nrules = 2, .zufd_cs_nrules = 2, .zufd_umc_style = ZEN_UMC_UMC_S_HYBRID_LPDDR5, .zufd_chan_hash = UMC_CHAN_HASH_F_BANK | UMC_CHAN_HASH_F_CS }, { .zufd_family = X86_PF_AMD_MENDOCINO, .zufd_flags = ZEN_UMC_FAM_F_NORM_HASH, .zufd_dram_nrules = 2, .zufd_cs_nrules = 2, .zufd_umc_style = ZEN_UMC_UMC_S_HYBRID_LPDDR5, .zufd_chan_hash = UMC_CHAN_HASH_F_BANK | UMC_CHAN_HASH_F_CS }, { .zufd_family = X86_PF_AMD_MILAN, .zufd_flags = ZEN_UMC_FAM_F_TARG_REMAP | ZEN_UMC_FAM_F_NP2 | ZEN_UMC_FAM_F_NORM_HASH | ZEN_UMC_FAM_F_UMC_HASH, .zufd_dram_nrules = 16, .zufd_cs_nrules = 2, .zufd_umc_style = ZEN_UMC_UMC_S_DDR4, .zufd_chan_hash = UMC_CHAN_HASH_F_BANK | UMC_CHAN_HASH_F_RM | UMC_CHAN_HASH_F_CS }, { .zufd_family = X86_PF_AMD_GENOA, .zufd_flags = ZEN_UMC_FAM_F_TARG_REMAP | ZEN_UMC_FAM_F_UMC_HASH | ZEN_UMC_FAM_F_UMC_EADDR | ZEN_UMC_FAM_F_CS_XOR, .zufd_dram_nrules = 20, .zufd_cs_nrules = 4, .zufd_umc_style = ZEN_UMC_UMC_S_DDR5, .zufd_chan_hash = UMC_CHAN_HASH_F_BANK | UMC_CHAN_HASH_F_RM | UMC_CHAN_HASH_F_PC | UMC_CHAN_HASH_F_CS }, { .zufd_family = X86_PF_AMD_VERMEER, .zufd_flags = ZEN_UMC_FAM_F_NORM_HASH | ZEN_UMC_FAM_F_UMC_HASH, .zufd_dram_nrules = 16, .zufd_cs_nrules = 2, .zufd_umc_style = ZEN_UMC_UMC_S_DDR4, .zufd_chan_hash = UMC_CHAN_HASH_F_BANK | UMC_CHAN_HASH_F_RM | UMC_CHAN_HASH_F_CS, }, { .zufd_family = X86_PF_AMD_REMBRANDT, .zufd_flags = ZEN_UMC_FAM_F_NORM_HASH, .zufd_dram_nrules = 2, .zufd_cs_nrules = 2, .zufd_umc_style = ZEN_UMC_UMC_S_DDR5_APU, .zufd_chan_hash = UMC_CHAN_HASH_F_BANK | UMC_CHAN_HASH_F_CS }, { .zufd_family = X86_PF_AMD_CEZANNE, .zufd_flags = ZEN_UMC_FAM_F_NORM_HASH, .zufd_dram_nrules = 2, .zufd_cs_nrules = 2, .zufd_umc_style = ZEN_UMC_UMC_S_DDR4_APU, .zufd_chan_hash = UMC_CHAN_HASH_F_BANK | UMC_CHAN_HASH_F_PC | UMC_CHAN_HASH_F_CS }, { .zufd_family = X86_PF_AMD_RAPHAEL, .zufd_flags = ZEN_UMC_FAM_F_TARG_REMAP | ZEN_UMC_FAM_F_CS_XOR, .zufd_dram_nrules = 2, .zufd_cs_nrules = 2, .zufd_umc_style = ZEN_UMC_UMC_S_DDR5, .zufd_chan_hash = UMC_CHAN_HASH_F_BANK | UMC_CHAN_HASH_F_PC | UMC_CHAN_HASH_F_CS }, { .zufd_family = X86_PF_AMD_PHOENIX, .zufd_flags = ZEN_UMC_FAM_F_CS_XOR, .zufd_dram_nrules = 2, .zufd_cs_nrules = 2, .zufd_umc_style = ZEN_UMC_UMC_S_DDR5_APU, .zufd_chan_hash = UMC_CHAN_HASH_F_BANK | UMC_CHAN_HASH_F_CS }, { .zufd_family = X86_PF_AMD_BERGAMO, .zufd_flags = ZEN_UMC_FAM_F_TARG_REMAP | ZEN_UMC_FAM_F_UMC_HASH | ZEN_UMC_FAM_F_UMC_EADDR | ZEN_UMC_FAM_F_CS_XOR, .zufd_dram_nrules = 20, .zufd_cs_nrules = 4, .zufd_umc_style = ZEN_UMC_UMC_S_DDR5, .zufd_chan_hash = UMC_CHAN_HASH_F_BANK | UMC_CHAN_HASH_F_RM | UMC_CHAN_HASH_F_PC | UMC_CHAN_HASH_F_CS } }; /* * We use this for the DDR4 and Hybrid DDR4 + LPDDR5 tables to map between the * specific enumerated speeds which are encoded values and the corresponding * memory clock and speed. For all DDR4 and LPDDR5 items we assume a a 1:2 ratio * between them. This is not used for the pure DDR5 / LPDDR5 entries because of * how the register just encodes the raw value in MHz. */ typedef struct zen_umc_freq_map { uint32_t zufm_reg; uint32_t zufm_mhz; uint32_t zufm_mts2; uint32_t zufm_mts4; } zen_umc_freq_map_t; static const zen_umc_freq_map_t zen_umc_ddr4_map[] = { { UMC_DRAMCFG_DDR4_MEMCLK_667, 667, 1333, 0 }, { UMC_DRAMCFG_DDR4_MEMCLK_800, 800, 1600, 0 }, { UMC_DRAMCFG_DDR4_MEMCLK_933, 933, 1866, 0 }, { UMC_DRAMCFG_DDR4_MEMCLK_1067, 1067, 2133, 0 }, { UMC_DRAMCFG_DDR4_MEMCLK_1200, 1200, 2400, 0 }, { UMC_DRAMCFG_DDR4_MEMCLK_1333, 1333, 2666, 0 }, { UMC_DRAMCFG_DDR4_MEMCLK_1467, 1467, 2933, 0 }, { UMC_DRAMCFG_DDR4_MEMCLK_1600, 1600, 3200, 0 } }; static const zen_umc_freq_map_t zen_umc_lpddr5_map[] = { { UMC_DRAMCFG_HYB_MEMCLK_333, 333, 667, 1333 }, { UMC_DRAMCFG_HYB_MEMCLK_400, 400, 800, 1600 }, { UMC_DRAMCFG_HYB_MEMCLK_533, 533, 1066, 2133 }, { UMC_DRAMCFG_HYB_MEMCLK_687, 687, 1375, 2750 }, { UMC_DRAMCFG_HYB_MEMCLK_750, 750, 1500, 3000 }, { UMC_DRAMCFG_HYB_MEMCLK_800, 800, 1600, 3200 }, { UMC_DRAMCFG_HYB_MEMCLK_933, 933, 1866, 3733 }, { UMC_DRAMCFG_HYB_MEMCLK_1066, 1066, 2133, 4267 }, { UMC_DRAMCFG_HYB_MEMCLK_1200, 1200, 2400, 4800 }, { UMC_DRAMCFG_HYB_MEMCLK_1375, 1375, 2750, 5500 }, { UMC_DRAMCFG_HYB_MEMCLK_1500, 1500, 3000, 6000 }, { UMC_DRAMCFG_HYB_MEMCLK_1600, 1600, 3200, 6400 } }; static boolean_t zen_umc_identify(zen_umc_t *umc) { for (uint_t i = 0; i < ARRAY_SIZE(zen_umc_fam_data); i++) { if (zen_umc_fam_data[i].zufd_family == umc->umc_family) { umc->umc_fdata = &zen_umc_fam_data[i]; return (B_TRUE); } } return (B_FALSE); } /* * This operates on DFv2, DFv3, and DFv3.5 DRAM rules, which generally speaking * are in similar register locations and meanings, but the size of bits in * memory is not consistent. */ static int zen_umc_read_dram_rule_df_23(zen_umc_t *umc, const uint_t dfno, const uint_t inst, const uint_t ruleno, df_dram_rule_t *rule) { int ret; uint32_t base, limit; uint64_t dbase, dlimit; uint16_t addr_ileave, chan_ileave, sock_ileave, die_ileave, dest; boolean_t hash = B_FALSE; zen_umc_df_t *df = &umc->umc_dfs[dfno]; if ((ret = amdzen_c_df_read32(dfno, inst, DF_DRAM_BASE_V2(ruleno), &base)) != 0) { dev_err(umc->umc_dip, CE_WARN, "!failed to read DRAM base " "register %u on 0x%x/0x%x: %d", ruleno, dfno, inst, ret); return (ret); } if ((ret = amdzen_c_df_read32(dfno, inst, DF_DRAM_LIMIT_V2(ruleno), &limit)) != 0) { dev_err(umc->umc_dip, CE_WARN, "!failed to read DRAM limit " "register %u on 0x%x/0x%x: %d", ruleno, dfno, inst, ret); return (ret); } rule->ddr_raw_base = base; rule->ddr_raw_limit = limit; rule->ddr_raw_ileave = rule->ddr_raw_ctrl = 0; if (!DF_DRAM_BASE_V2_GET_VALID(base)) { return (0); } /* * Extract all values from the registers and then normalize. While there * are often different bit patterns for the values, the interpretation * is the same across all the Zen 1-3 parts. That is while which bits * may be used for say channel interleave vary, the values of them are * consistent. */ rule->ddr_flags |= DF_DRAM_F_VALID; if (DF_DRAM_BASE_V2_GET_HOLE_EN(base)) { rule->ddr_flags |= DF_DRAM_F_HOLE; } dbase = DF_DRAM_BASE_V2_GET_BASE(base); dlimit = DF_DRAM_LIMIT_V2_GET_LIMIT(limit); switch (umc->umc_df_rev) { case DF_REV_2: addr_ileave = DF_DRAM_BASE_V2_GET_ILV_ADDR(base); chan_ileave = DF_DRAM_BASE_V2_GET_ILV_CHAN(base); die_ileave = DF_DRAM_LIMIT_V2_GET_ILV_DIE(limit); sock_ileave = DF_DRAM_LIMIT_V2_GET_ILV_SOCK(limit); dest = DF_DRAM_LIMIT_V2_GET_DEST_ID(limit); break; case DF_REV_3: addr_ileave = DF_DRAM_BASE_V3_GET_ILV_ADDR(base); sock_ileave = DF_DRAM_BASE_V3_GET_ILV_SOCK(base); die_ileave = DF_DRAM_BASE_V3_GET_ILV_DIE(base); chan_ileave = DF_DRAM_BASE_V3_GET_ILV_CHAN(base); dest = DF_DRAM_LIMIT_V3_GET_DEST_ID(limit); break; case DF_REV_3P5: addr_ileave = DF_DRAM_BASE_V3P5_GET_ILV_ADDR(base); sock_ileave = DF_DRAM_BASE_V3P5_GET_ILV_SOCK(base); die_ileave = DF_DRAM_BASE_V3P5_GET_ILV_DIE(base); chan_ileave = DF_DRAM_BASE_V3P5_GET_ILV_CHAN(base); dest = DF_DRAM_LIMIT_V3P5_GET_DEST_ID(limit); break; default: dev_err(umc->umc_dip, CE_WARN, "!encountered unsupported " "DF revision processing DRAM rules: 0x%x", umc->umc_df_rev); return (-1); } rule->ddr_base = dbase << DF_DRAM_BASE_V2_BASE_SHIFT; rule->ddr_sock_ileave_bits = sock_ileave; rule->ddr_die_ileave_bits = die_ileave; switch (addr_ileave) { case DF_DRAM_ILV_ADDR_8: case DF_DRAM_ILV_ADDR_9: case DF_DRAM_ILV_ADDR_10: case DF_DRAM_ILV_ADDR_11: case DF_DRAM_ILV_ADDR_12: break; default: dev_err(umc->umc_dip, CE_WARN, "!encountered invalid address " "interleave on rule %u, df/inst 0x%x/0x%x: 0x%x", ruleno, dfno, inst, addr_ileave); return (EINVAL); } rule->ddr_addr_start = DF_DRAM_ILV_ADDR_BASE + addr_ileave; switch (chan_ileave) { case DF_DRAM_BASE_V2_ILV_CHAN_1: rule->ddr_chan_ileave = DF_CHAN_ILEAVE_1CH; break; case DF_DRAM_BASE_V2_ILV_CHAN_2: rule->ddr_chan_ileave = DF_CHAN_ILEAVE_2CH; break; case DF_DRAM_BASE_V2_ILV_CHAN_4: rule->ddr_chan_ileave = DF_CHAN_ILEAVE_4CH; break; case DF_DRAM_BASE_V2_ILV_CHAN_8: rule->ddr_chan_ileave = DF_CHAN_ILEAVE_8CH; break; case DF_DRAM_BASE_V2_ILV_CHAN_6: rule->ddr_chan_ileave = DF_CHAN_ILEAVE_6CH; break; case DF_DRAM_BASE_V2_ILV_CHAN_COD4_2: hash = B_TRUE; rule->ddr_chan_ileave = DF_CHAN_ILEAVE_COD4_2CH; break; case DF_DRAM_BASE_V2_ILV_CHAN_COD2_4: hash = B_TRUE; rule->ddr_chan_ileave = DF_CHAN_ILEAVE_COD2_4CH; break; case DF_DRAM_BASE_V2_ILV_CHAN_COD1_8: hash = B_TRUE; rule->ddr_chan_ileave = DF_CHAN_ILEAVE_COD1_8CH; break; default: dev_err(umc->umc_dip, CE_WARN, "!encountered invalid channel " "interleave on rule %u, df/inst 0x%x/0x%x: 0x%x", ruleno, dfno, inst, chan_ileave); return (EINVAL); } /* * If hashing is enabled, note which hashing rules apply to this * address. This is done to smooth over the differences between DFv3 and * DFv4, where the flags are in the rules themselves in the latter, but * global today. */ if (hash) { if ((df->zud_flags & ZEN_UMC_DF_F_HASH_16_18) != 0) { rule->ddr_flags |= DF_DRAM_F_HASH_16_18; } if ((df->zud_flags & ZEN_UMC_DF_F_HASH_21_23) != 0) { rule->ddr_flags |= DF_DRAM_F_HASH_21_23; } if ((df->zud_flags & ZEN_UMC_DF_F_HASH_30_32) != 0) { rule->ddr_flags |= DF_DRAM_F_HASH_30_32; } } /* * While DFv4 makes remapping explicit, it is basically always enabled * and used on supported platforms prior to that point. So flag such * supported platforms as ones that need to do this. On those systems * there is only one set of remap rules for an entire DF that are * determined based on the target socket. To indicate that we use the * DF_DRAM_F_REMAP_SOCK flag below and skip setting a remap target. */ if ((umc->umc_fdata->zufd_flags & ZEN_UMC_FAM_F_TARG_REMAP) != 0) { rule->ddr_flags |= DF_DRAM_F_REMAP_EN | DF_DRAM_F_REMAP_SOCK; } rule->ddr_limit = (dlimit << DF_DRAM_LIMIT_V2_LIMIT_SHIFT) + DF_DRAM_LIMIT_V2_LIMIT_EXCL; rule->ddr_dest_fabid = dest; return (0); } static int zen_umc_read_dram_rule_df_4(zen_umc_t *umc, const uint_t dfno, const uint_t inst, const uint_t ruleno, df_dram_rule_t *rule) { int ret; uint16_t addr_ileave; uint32_t base, limit, ilv, ctl; if ((ret = amdzen_c_df_read32(dfno, inst, DF_DRAM_BASE_V4(ruleno), &base)) != 0) { dev_err(umc->umc_dip, CE_WARN, "!failed to read DRAM base " "register %u on 0x%x/0x%x: %d", ruleno, dfno, inst, ret); return (ret); } if ((ret = amdzen_c_df_read32(dfno, inst, DF_DRAM_LIMIT_V4(ruleno), &limit)) != 0) { dev_err(umc->umc_dip, CE_WARN, "!failed to read DRAM limit " "register %u on 0x%x/0x%x: %d", ruleno, dfno, inst, ret); return (ret); } if ((ret = amdzen_c_df_read32(dfno, inst, DF_DRAM_ILV_V4(ruleno), &ilv)) != 0) { dev_err(umc->umc_dip, CE_WARN, "!failed to read DRAM " "interleave register %u on 0x%x/0x%x: %d", ruleno, dfno, inst, ret); return (ret); } if ((ret = amdzen_c_df_read32(dfno, inst, DF_DRAM_CTL_V4(ruleno), &ctl)) != 0) { dev_err(umc->umc_dip, CE_WARN, "!failed to read DRAM control " "register %u on 0x%x/0x%x: %d", ruleno, dfno, inst, ret); return (ret); } rule->ddr_raw_base = base; rule->ddr_raw_limit = limit; rule->ddr_raw_ileave = ilv; rule->ddr_raw_ctrl = ctl; if (!DF_DRAM_CTL_V4_GET_VALID(ctl)) { return (0); } rule->ddr_flags |= DF_DRAM_F_VALID; rule->ddr_base = DF_DRAM_BASE_V4_GET_ADDR(base); rule->ddr_base = rule->ddr_base << DF_DRAM_BASE_V4_BASE_SHIFT; rule->ddr_limit = DF_DRAM_LIMIT_V4_GET_ADDR(limit); rule->ddr_limit = (rule->ddr_limit << DF_DRAM_LIMIT_V4_LIMIT_SHIFT) + DF_DRAM_LIMIT_V4_LIMIT_EXCL; rule->ddr_dest_fabid = DF_DRAM_CTL_V4_GET_DEST_ID(ctl); if (DF_DRAM_CTL_V4_GET_HASH_1G(ctl) != 0) { rule->ddr_flags |= DF_DRAM_F_HASH_30_32; } if (DF_DRAM_CTL_V4_GET_HASH_2M(ctl) != 0) { rule->ddr_flags |= DF_DRAM_F_HASH_21_23; } if (DF_DRAM_CTL_V4_GET_HASH_64K(ctl) != 0) { rule->ddr_flags |= DF_DRAM_F_HASH_16_18; } if (DF_DRAM_CTL_V4_GET_REMAP_EN(ctl) != 0) { rule->ddr_flags |= DF_DRAM_F_REMAP_EN; rule->ddr_remap_ent = DF_DRAM_CTL_V4_GET_REMAP_SEL(ctl); } if (DF_DRAM_CTL_V4_GET_HOLE_EN(ctl) != 0) { rule->ddr_flags |= DF_DRAM_F_HOLE; } rule->ddr_sock_ileave_bits = DF_DRAM_ILV_V4_GET_SOCK(ilv); rule->ddr_die_ileave_bits = DF_DRAM_ILV_V4_GET_DIE(ilv); switch (DF_DRAM_ILV_V4_GET_CHAN(ilv)) { case DF_DRAM_ILV_V4_CHAN_1: rule->ddr_chan_ileave = DF_CHAN_ILEAVE_1CH; break; case DF_DRAM_ILV_V4_CHAN_2: rule->ddr_chan_ileave = DF_CHAN_ILEAVE_2CH; break; case DF_DRAM_ILV_V4_CHAN_4: rule->ddr_chan_ileave = DF_CHAN_ILEAVE_4CH; break; case DF_DRAM_ILV_V4_CHAN_8: rule->ddr_chan_ileave = DF_CHAN_ILEAVE_8CH; break; case DF_DRAM_ILV_V4_CHAN_16: rule->ddr_chan_ileave = DF_CHAN_ILEAVE_16CH; break; case DF_DRAM_ILV_V4_CHAN_32: rule->ddr_chan_ileave = DF_CHAN_ILEAVE_32CH; break; case DF_DRAM_ILV_V4_CHAN_NPS4_2CH: rule->ddr_chan_ileave = DF_CHAN_ILEAVE_NPS4_2CH; break; case DF_DRAM_ILV_V4_CHAN_NPS2_4CH: rule->ddr_chan_ileave = DF_CHAN_ILEAVE_COD2_4CH; break; case DF_DRAM_ILV_V4_CHAN_NPS1_8CH: rule->ddr_chan_ileave = DF_CHAN_ILEAVE_NPS1_8CH; break; case DF_DRAM_ILV_V4_CHAN_NPS4_3CH: rule->ddr_chan_ileave = DF_CHAN_ILEAVE_NPS4_3CH; break; case DF_DRAM_ILV_V4_CHAN_NPS2_6CH: rule->ddr_chan_ileave = DF_CHAN_ILEAVE_NPS2_6CH; break; case DF_DRAM_ILV_V4_CHAN_NPS1_12CH: rule->ddr_chan_ileave = DF_CHAN_ILEAVE_NPS1_12CH; break; case DF_DRAM_ILV_V4_CHAN_NPS2_5CH: rule->ddr_chan_ileave = DF_CHAN_ILEAVE_NPS2_5CH; break; case DF_DRAM_ILV_V4_CHAN_NPS1_10CH: rule->ddr_chan_ileave = DF_CHAN_ILEAVE_NPS1_10CH; break; default: dev_err(umc->umc_dip, CE_WARN, "!encountered invalid channel " "interleave on rule %u, df/inst 0x%x/0x%x: 0x%x", ruleno, dfno, inst, DF_DRAM_ILV_V4_GET_CHAN(ilv)); break; } addr_ileave = DF_DRAM_ILV_V4_GET_ADDR(ilv); switch (addr_ileave) { case DF_DRAM_ILV_ADDR_8: case DF_DRAM_ILV_ADDR_9: case DF_DRAM_ILV_ADDR_10: case DF_DRAM_ILV_ADDR_11: case DF_DRAM_ILV_ADDR_12: break; default: dev_err(umc->umc_dip, CE_WARN, "!encountered invalid address " "interleave on rule %u, df/inst 0x%x/0x%x: 0x%x", ruleno, dfno, inst, addr_ileave); return (EINVAL); } rule->ddr_addr_start = DF_DRAM_ILV_ADDR_BASE + addr_ileave; return (0); } static int zen_umc_read_dram_rule(zen_umc_t *umc, const uint_t dfno, const uint_t instid, const uint_t ruleno, df_dram_rule_t *rule) { int ret; switch (umc->umc_df_rev) { case DF_REV_2: case DF_REV_3: case DF_REV_3P5: ret = zen_umc_read_dram_rule_df_23(umc, dfno, instid, ruleno, rule); break; case DF_REV_4: ret = zen_umc_read_dram_rule_df_4(umc, dfno, instid, ruleno, rule); break; default: dev_err(umc->umc_dip, CE_WARN, "!encountered unsupported " "DF revision processing DRAM rules: 0x%x", umc->umc_df_rev); return (-1); } if (ret != 0) { dev_err(umc->umc_dip, CE_WARN, "!failed to read DRAM " "rule %u on df/inst 0x%x/0x%x: %d", ruleno, dfno, instid, ret); return (-1); } return (0); } static int zen_umc_read_remap(zen_umc_t *umc, zen_umc_df_t *df, const uint_t instid) { uint_t nremaps, nents; uint_t dfno = df->zud_dfno; const df_reg_def_t milan_remap0[ZEN_UMC_MILAN_CS_NREMAPS] = { DF_SKT0_CS_REMAP0_V3, DF_SKT1_CS_REMAP0_V3 }; const df_reg_def_t milan_remap1[ZEN_UMC_MILAN_CS_NREMAPS] = { DF_SKT0_CS_REMAP1_V3, DF_SKT1_CS_REMAP1_V3 }; const df_reg_def_t dfv4_remapA[ZEN_UMC_MAX_CS_REMAPS] = { DF_CS_REMAP0A_V4, DF_CS_REMAP1A_V4, DF_CS_REMAP2A_V4, DF_CS_REMAP3A_V4 }; const df_reg_def_t dfv4_remapB[ZEN_UMC_MAX_CS_REMAPS] = { DF_CS_REMAP0B_V4, DF_CS_REMAP1B_V4, DF_CS_REMAP2B_V4, DF_CS_REMAP3B_V4 }; const df_reg_def_t *remapA, *remapB; switch (umc->umc_df_rev) { case DF_REV_3: nremaps = ZEN_UMC_MILAN_CS_NREMAPS; nents = ZEN_UMC_MILAN_REMAP_ENTS; remapA = milan_remap0; remapB = milan_remap1; break; case DF_REV_4: nremaps = ZEN_UMC_MAX_CS_REMAPS; nents = ZEN_UMC_MAX_REMAP_ENTS; remapA = dfv4_remapA; remapB = dfv4_remapB; break; default: dev_err(umc->umc_dip, CE_WARN, "!encountered unsupported DF " "revision processing remap rules: 0x%x", umc->umc_df_rev); return (-1); } df->zud_cs_nremap = nremaps; for (uint_t i = 0; i < nremaps; i++) { int ret; uint32_t rmA, rmB; zen_umc_cs_remap_t *remap = &df->zud_remap[i]; if ((ret = amdzen_c_df_read32(dfno, instid, remapA[i], &rmA)) != 0) { dev_err(umc->umc_dip, CE_WARN, "!failed to read " "df/inst 0x%x/0x%x remap socket %u-0/A: %d", dfno, instid, i, ret); return (-1); } if ((ret = amdzen_c_df_read32(dfno, instid, remapB[i], &rmB)) != 0) { dev_err(umc->umc_dip, CE_WARN, "!failed to read " "df/inst 0x%x/0x%x remap socket %u-1/B: %d", dfno, instid, i, ret); return (-1); } remap->csr_nremaps = nents; for (uint_t ent = 0; ent < ZEN_UMC_REMAP_PER_REG; ent++) { uint_t alt = ent + ZEN_UMC_REMAP_PER_REG; boolean_t do_alt = alt < nents; remap->csr_remaps[ent] = DF_CS_REMAP_GET_CSX(rmA, ent); if (do_alt) { remap->csr_remaps[alt] = DF_CS_REMAP_GET_CSX(rmB, ent); } } } return (0); } /* * Now that we have a CCM, we have several different tasks ahead of us: * * o Determine whether or not the DRAM hole is valid. * o Snapshot all of the system address rules and translate them into our * generic format. * o Determine if there are any rules to retarget things (currently * Milan/Genoa). * o Determine if there are any other hashing rules enabled. * * We only require this from a single CCM as these are currently required to be * the same across all of them. */ static int zen_umc_fill_ccm_cb(const uint_t dfno, const uint32_t fabid, const uint32_t instid, void *arg) { zen_umc_t *umc = arg; zen_umc_df_t *df = &umc->umc_dfs[dfno]; df_reg_def_t hole; int ret; uint32_t val; df->zud_dfno = dfno; df->zud_ccm_inst = instid; /* * First get the DRAM hole. This has the same layout, albeit different * registers across our different platforms. */ switch (umc->umc_df_rev) { case DF_REV_2: case DF_REV_3: case DF_REV_3P5: hole = DF_DRAM_HOLE_V2; break; case DF_REV_4: hole = DF_DRAM_HOLE_V4; break; default: dev_err(umc->umc_dip, CE_WARN, "!encountered unsupported " "DF version: 0x%x", umc->umc_df_rev); return (-1); } if ((ret = amdzen_c_df_read32(dfno, instid, hole, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "!failed to read DRAM Hole: %d", ret); return (-1); } df->zud_hole_raw = val; if (DF_DRAM_HOLE_GET_VALID(val)) { uint64_t t; df->zud_flags |= ZEN_UMC_DF_F_HOLE_VALID; t = DF_DRAM_HOLE_GET_BASE(val); df->zud_hole_base = t << DF_DRAM_HOLE_BASE_SHIFT; } /* * Prior to Zen 4, the hash information was global and applied to all * COD rules globally. Check if we're on such a system and snapshot this * so we can use it during the rule application. Note, this was added in * DFv3. */ if (umc->umc_df_rev == DF_REV_3 || umc->umc_df_rev == DF_REV_3P5) { uint32_t globctl; if ((ret = amdzen_c_df_read32(dfno, instid, DF_GLOB_CTL_V3, &globctl)) != 0) { dev_err(umc->umc_dip, CE_WARN, "!failed to read global " "control: %d", ret); return (-1); } df->zud_glob_ctl_raw = globctl; if (DF_GLOB_CTL_V3_GET_HASH_1G(globctl) != 0) { df->zud_flags |= ZEN_UMC_DF_F_HASH_30_32; } if (DF_GLOB_CTL_V3_GET_HASH_2M(globctl) != 0) { df->zud_flags |= ZEN_UMC_DF_F_HASH_21_23; } if (DF_GLOB_CTL_V3_GET_HASH_64K(globctl) != 0) { df->zud_flags |= ZEN_UMC_DF_F_HASH_16_18; } } df->zud_dram_nrules = umc->umc_fdata->zufd_dram_nrules; for (uint_t i = 0; i < umc->umc_fdata->zufd_dram_nrules; i++) { if (zen_umc_read_dram_rule(umc, dfno, instid, i, &df->zud_rules[i]) != 0) { return (-1); } } if ((umc->umc_fdata->zufd_flags & ZEN_UMC_FAM_F_TARG_REMAP) != 0) { if (zen_umc_read_remap(umc, df, instid) != 0) { return (-1); } } /* * We only want a single entry, so always return 1 to terminate us * early. */ return (1); } /* * At this point we can go through and calculate the size of the DIMM that we've * found. While it would be nice to determine this from the SPD data, we can * figure this out entirely based upon the information in the memory controller. * * This works by first noting that DDR4, LPDDR4, DDR5, and LPDDR5 are all built * around 64-bit data channels. This means that each row and column provides up * 64-bits (ignoring ECC) of data. There are a number of banks and bank groups. * The memory controller tracks the total number of bits that are used for each. * While DDR5 introduces sub-channels, we don't need to worry about those here, * because ultimately the sub-channel just splits the 64-bit bus we're assuming * into 2x 32-bit buses. While they can be independently selected, they should * have equivalent capacities. * * The most confusing part of this is that there is one of these related to each * rank on the device. The UMC natively has two 'chip-selects', each of which is * used to correspond to a rank. There are then separately multiple rm bits in * each chip-select. As far as we can tell the PSP or SMU programs the number of * rm bits to be zero when you have a dual-rank device. * * We end up summing each chip-select rather than assuming that the chip-selects * are identical. In theory some amount of asymmetric DIMMs exist in the wild, * but we don't know of many systems using them. */ static void zen_umc_calc_dimm_size(umc_dimm_t *dimm) { dimm->ud_dimm_size = 0; for (uint_t i = 0; i < ZEN_UMC_MAX_CHAN_BASE; i++) { uint64_t nrc; const umc_cs_t *cs = &dimm->ud_cs[i]; if (!cs->ucs_base.udb_valid && !cs->ucs_sec.udb_valid) { continue; } nrc = cs->ucs_nrow_lo + cs->ucs_nrow_hi + cs->ucs_ncol; dimm->ud_dimm_size += (8ULL << nrc) * (1 << cs->ucs_nbanks) * (1 << cs->ucs_nrm); } } /* * This is used to fill in the common properties about a DIMM. This should occur * after the rank information has been filled out. The information used is the * same between DDR4 and DDR5 DIMMs. The only major difference is the register * offset. */ static boolean_t zen_umc_fill_dimm_common(zen_umc_t *umc, zen_umc_df_t *df, zen_umc_chan_t *chan, const uint_t dimmno, boolean_t ddr4_style) { umc_dimm_t *dimm; int ret; smn_reg_t reg; uint32_t val; const uint32_t id = chan->chan_logid; dimm = &chan->chan_dimms[dimmno]; dimm->ud_dimmno = dimmno; if (ddr4_style) { reg = UMC_DIMMCFG_DDR4(id, dimmno); } else { reg = UMC_DIMMCFG_DDR5(id, dimmno); } if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read DIMM " "configuration register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } dimm->ud_dimmcfg_raw = val; if (UMC_DIMMCFG_GET_X16(val) != 0) { dimm->ud_width = UMC_DIMM_W_X16; } else if (UMC_DIMMCFG_GET_X4(val) != 0) { dimm->ud_width = UMC_DIMM_W_X4; } else { dimm->ud_width = UMC_DIMM_W_X8; } if (UMC_DIMMCFG_GET_3DS(val) != 0) { dimm->ud_kind = UMC_DIMM_K_3DS_RDIMM; } else if (UMC_DIMMCFG_GET_LRDIMM(val) != 0) { dimm->ud_kind = UMC_DIMM_K_LRDIMM; } else if (UMC_DIMMCFG_GET_RDIMM(val) != 0) { dimm->ud_kind = UMC_DIMM_K_RDIMM; } else { dimm->ud_kind = UMC_DIMM_K_UDIMM; } /* * DIMM information in a UMC can be somewhat confusing. There are quite * a number of non-zero reset values that are here. Flag whether or not * we think this entry should be usable based on enabled chip-selects. */ for (uint_t i = 0; i < ZEN_UMC_MAX_CHAN_BASE; i++) { if (dimm->ud_cs[i].ucs_base.udb_valid || dimm->ud_cs[i].ucs_sec.udb_valid) { dimm->ud_flags |= UMC_DIMM_F_VALID; break; } } /* * The remaining calculations we only want to perform if we have actual * data for a DIMM. */ if ((dimm->ud_flags & UMC_DIMM_F_VALID) == 0) { return (B_TRUE); } zen_umc_calc_dimm_size(dimm); return (B_TRUE); } /* * Fill all the information about a DDR4 DIMM. In the DDR4 UMC, some of this * information is on a per-chip select basis while at other times it is on a * per-DIMM basis. In general, chip-selects 0/1 correspond to DIMM 0, and * chip-selects 2/3 correspond to DIMM 1. To normalize things with the DDR5 UMC * which generally has things stored on a per-rank/chips-select basis, we * duplicate information that is DIMM-wide into the chip-select data structure * (umc_cs_t). */ static boolean_t zen_umc_fill_chan_dimm_ddr4(zen_umc_t *umc, zen_umc_df_t *df, zen_umc_chan_t *chan, const uint_t dimmno) { umc_dimm_t *dimm; umc_cs_t *cs0, *cs1; const uint32_t id = chan->chan_logid; int ret; uint32_t val; smn_reg_t reg; ASSERT3U(dimmno, <, ZEN_UMC_MAX_DIMMS); dimm = &chan->chan_dimms[dimmno]; cs0 = &dimm->ud_cs[0]; cs1 = &dimm->ud_cs[1]; /* * DDR4 organization has initial data that exists on a per-chip select * basis. The rest of it is on a per-DIMM basis. First we grab the * per-chip-select data. After this for loop, we will always duplicate * all data that we gather into both chip-selects. */ for (uint_t i = 0; i < ZEN_UMC_MAX_CS_PER_DIMM; i++) { uint64_t addr; const uint16_t reginst = i + dimmno * 2; reg = UMC_BASE(id, reginst); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read base " "register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } addr = (uint64_t)UMC_BASE_GET_ADDR(val) << UMC_BASE_ADDR_SHIFT; dimm->ud_cs[i].ucs_base.udb_base = addr; dimm->ud_cs[i].ucs_base.udb_valid = UMC_BASE_GET_EN(val); reg = UMC_BASE_SEC(id, reginst); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read " "secondary base register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } addr = (uint64_t)UMC_BASE_GET_ADDR(val) << UMC_BASE_ADDR_SHIFT; dimm->ud_cs[i].ucs_sec.udb_base = addr; dimm->ud_cs[i].ucs_sec.udb_valid = UMC_BASE_GET_EN(val); } reg = UMC_MASK_DDR4(id, dimmno); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read mask register " "%x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } /* * When we extract the masks, hardware only checks a limited range of * bits. Therefore we need to always OR in those lower order bits. */ cs0->ucs_base_mask = (uint64_t)UMC_MASK_GET_ADDR(val) << UMC_MASK_ADDR_SHIFT; cs0->ucs_base_mask |= (1 << UMC_MASK_ADDR_SHIFT) - 1; cs1->ucs_base_mask = cs0->ucs_base_mask; reg = UMC_MASK_SEC_DDR4(id, dimmno); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read secondary mask " "register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } cs0->ucs_sec_mask = (uint64_t)UMC_MASK_GET_ADDR(val) << UMC_MASK_ADDR_SHIFT; cs0->ucs_sec_mask |= (1 << UMC_MASK_ADDR_SHIFT) - 1; cs1->ucs_sec_mask = cs0->ucs_sec_mask; reg = UMC_ADDRCFG_DDR4(id, dimmno); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read address config " "register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } cs0->ucs_nbanks = UMC_ADDRCFG_GET_NBANK_BITS(val) + UMC_ADDRCFG_NBANK_BITS_BASE; cs1->ucs_nbanks = cs0->ucs_nbanks; cs0->ucs_ncol = UMC_ADDRCFG_GET_NCOL_BITS(val) + UMC_ADDRCFG_NCOL_BITS_BASE; cs1->ucs_ncol = cs0->ucs_ncol; cs0->ucs_nrow_hi = UMC_ADDRCFG_DDR4_GET_NROW_BITS_HI(val); cs1->ucs_nrow_hi = cs0->ucs_nrow_hi; cs0->ucs_nrow_lo = UMC_ADDRCFG_GET_NROW_BITS_LO(val) + UMC_ADDRCFG_NROW_BITS_LO_BASE; cs1->ucs_nrow_lo = cs0->ucs_nrow_lo; cs0->ucs_nbank_groups = UMC_ADDRCFG_GET_NBANKGRP_BITS(val); cs1->ucs_nbank_groups = cs0->ucs_nbank_groups; /* * As the chip-select XORs don't always show up, use a dummy value * that'll result in no change occurring here. */ cs0->ucs_cs_xor = cs1->ucs_cs_xor = 0; /* * APUs don't seem to support various rank select bits. */ if (umc->umc_fdata->zufd_umc_style == ZEN_UMC_UMC_S_DDR4) { cs0->ucs_nrm = UMC_ADDRCFG_DDR4_GET_NRM_BITS(val); cs1->ucs_nrm = cs0->ucs_nrm; } else { cs0->ucs_nrm = cs1->ucs_nrm = 0; } reg = UMC_ADDRSEL_DDR4(id, dimmno); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read bank address " "select register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } cs0->ucs_row_hi_bit = UMC_ADDRSEL_DDR4_GET_ROW_HI(val) + UMC_ADDRSEL_DDR4_ROW_HI_BASE; cs1->ucs_row_hi_bit = cs0->ucs_row_hi_bit; cs0->ucs_row_low_bit = UMC_ADDRSEL_GET_ROW_LO(val) + UMC_ADDRSEL_ROW_LO_BASE; cs1->ucs_row_low_bit = cs0->ucs_row_low_bit; cs0->ucs_bank_bits[0] = UMC_ADDRSEL_GET_BANK0(val) + UMC_ADDRSEL_BANK_BASE; cs0->ucs_bank_bits[1] = UMC_ADDRSEL_GET_BANK1(val) + UMC_ADDRSEL_BANK_BASE; cs0->ucs_bank_bits[2] = UMC_ADDRSEL_GET_BANK2(val) + UMC_ADDRSEL_BANK_BASE; cs0->ucs_bank_bits[3] = UMC_ADDRSEL_GET_BANK3(val) + UMC_ADDRSEL_BANK_BASE; cs0->ucs_bank_bits[4] = UMC_ADDRSEL_GET_BANK4(val) + UMC_ADDRSEL_BANK_BASE; bcopy(cs0->ucs_bank_bits, cs1->ucs_bank_bits, sizeof (cs0->ucs_bank_bits)); reg = UMC_COLSEL_LO_DDR4(id, dimmno); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read column address " "select low register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } for (uint_t i = 0; i < ZEN_UMC_MAX_COLSEL_PER_REG; i++) { cs0->ucs_col_bits[i] = UMC_COLSEL_REMAP_GET_COL(val, i) + UMC_COLSEL_LO_BASE; } reg = UMC_COLSEL_HI_DDR4(id, dimmno); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read column address " "select high register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } for (uint_t i = 0; i < ZEN_UMC_MAX_COLSEL_PER_REG; i++) { cs0->ucs_col_bits[i + ZEN_UMC_MAX_COLSEL_PER_REG] = UMC_COLSEL_REMAP_GET_COL(val, i) + UMC_COLSEL_HI_BASE; } bcopy(cs0->ucs_col_bits, cs1->ucs_col_bits, sizeof (cs0->ucs_col_bits)); /* * The next two registers give us information about a given rank select. * In the APUs, the inversion bits are there; however, the actual bit * selects are not. In this case we read the reserved bits regardless. * They should be ignored due to the fact that the number of banks is * zero. */ reg = UMC_RMSEL_DDR4(id, dimmno); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read rank address " "select register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } cs0->ucs_inv_msbs = UMC_RMSEL_DDR4_GET_INV_MSBE(val); cs1->ucs_inv_msbs = UMC_RMSEL_DDR4_GET_INV_MSBO(val); cs0->ucs_rm_bits[0] = UMC_RMSEL_DDR4_GET_RM0(val) + UMC_RMSEL_BASE; cs0->ucs_rm_bits[1] = UMC_RMSEL_DDR4_GET_RM1(val) + UMC_RMSEL_BASE; cs0->ucs_rm_bits[2] = UMC_RMSEL_DDR4_GET_RM2(val) + UMC_RMSEL_BASE; bcopy(cs0->ucs_rm_bits, cs1->ucs_rm_bits, sizeof (cs0->ucs_rm_bits)); reg = UMC_RMSEL_SEC_DDR4(id, dimmno); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read secondary rank " "address select register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } cs0->ucs_inv_msbs_sec = UMC_RMSEL_DDR4_GET_INV_MSBE(val); cs1->ucs_inv_msbs_sec = UMC_RMSEL_DDR4_GET_INV_MSBO(val); cs0->ucs_rm_bits_sec[0] = UMC_RMSEL_DDR4_GET_RM0(val) + UMC_RMSEL_BASE; cs0->ucs_rm_bits_sec[1] = UMC_RMSEL_DDR4_GET_RM1(val) + UMC_RMSEL_BASE; cs0->ucs_rm_bits_sec[2] = UMC_RMSEL_DDR4_GET_RM2(val) + UMC_RMSEL_BASE; bcopy(cs0->ucs_rm_bits_sec, cs1->ucs_rm_bits_sec, sizeof (cs0->ucs_rm_bits_sec)); return (zen_umc_fill_dimm_common(umc, df, chan, dimmno, B_TRUE)); } /* * The DDR5 based systems are organized such that almost all the information we * care about is split between two different chip-select structures in the UMC * hardware SMN space. */ static boolean_t zen_umc_fill_chan_rank_ddr5(zen_umc_t *umc, zen_umc_df_t *df, zen_umc_chan_t *chan, const uint_t dimmno, const uint_t rankno) { int ret; umc_cs_t *cs; uint32_t val; smn_reg_t reg; const uint32_t id = chan->chan_logid; const uint32_t regno = dimmno * 2 + rankno; ASSERT3U(dimmno, <, ZEN_UMC_MAX_DIMMS); ASSERT3U(rankno, <, ZEN_UMC_MAX_CS_PER_DIMM); cs = &chan->chan_dimms[dimmno].ud_cs[rankno]; reg = UMC_BASE(id, regno); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read base " "register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } cs->ucs_base.udb_base = (uint64_t)UMC_BASE_GET_ADDR(val) << UMC_BASE_ADDR_SHIFT; cs->ucs_base.udb_valid = UMC_BASE_GET_EN(val); if ((umc->umc_fdata->zufd_flags & ZEN_UMC_FAM_F_UMC_EADDR) != 0) { uint64_t addr; reg = UMC_BASE_EXT_DDR5(id, regno); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read " "extended base register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } addr = (uint64_t)UMC_BASE_EXT_GET_ADDR(val) << UMC_BASE_EXT_ADDR_SHIFT; cs->ucs_base.udb_base |= addr; } reg = UMC_BASE_SEC(id, regno); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read secondary base " "register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } cs->ucs_sec.udb_base = (uint64_t)UMC_BASE_GET_ADDR(val) << UMC_BASE_ADDR_SHIFT; cs->ucs_sec.udb_valid = UMC_BASE_GET_EN(val); if ((umc->umc_fdata->zufd_flags & ZEN_UMC_FAM_F_UMC_EADDR) != 0) { uint64_t addr; reg = UMC_BASE_EXT_SEC_DDR5(id, regno); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read " "extended secondary base register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } addr = (uint64_t)UMC_BASE_EXT_GET_ADDR(val) << UMC_BASE_EXT_ADDR_SHIFT; cs->ucs_sec.udb_base |= addr; } reg = UMC_MASK_DDR5(id, regno); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read mask " "register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } cs->ucs_base_mask = (uint64_t)UMC_MASK_GET_ADDR(val) << UMC_MASK_ADDR_SHIFT; cs->ucs_base_mask |= (1 << UMC_MASK_ADDR_SHIFT) - 1; if ((umc->umc_fdata->zufd_flags & ZEN_UMC_FAM_F_UMC_EADDR) != 0) { uint64_t addr; reg = UMC_MASK_EXT_DDR5(id, regno); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read " "extended mask register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } addr = (uint64_t)UMC_MASK_EXT_GET_ADDR(val) << UMC_MASK_EXT_ADDR_SHIFT; cs->ucs_base_mask |= addr; } reg = UMC_MASK_SEC_DDR5(id, regno); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read secondary mask " "register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } cs->ucs_sec_mask = (uint64_t)UMC_MASK_GET_ADDR(val) << UMC_MASK_ADDR_SHIFT; cs->ucs_sec_mask |= (1 << UMC_MASK_ADDR_SHIFT) - 1; if ((umc->umc_fdata->zufd_flags & ZEN_UMC_FAM_F_UMC_EADDR) != 0) { uint64_t addr; reg = UMC_MASK_EXT_SEC_DDR5(id, regno); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read " "extended mask register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } addr = (uint64_t)UMC_MASK_EXT_GET_ADDR(val) << UMC_MASK_EXT_ADDR_SHIFT; cs->ucs_sec_mask |= addr; } reg = UMC_ADDRCFG_DDR5(id, regno); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read address config " "register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } if ((umc->umc_fdata->zufd_flags & ZEN_UMC_FAM_F_CS_XOR) != 0) { cs->ucs_cs_xor = UMC_ADDRCFG_DDR5_GET_CSXOR(val); } else { cs->ucs_cs_xor = 0; } cs->ucs_nbanks = UMC_ADDRCFG_GET_NBANK_BITS(val) + UMC_ADDRCFG_NBANK_BITS_BASE; cs->ucs_ncol = UMC_ADDRCFG_GET_NCOL_BITS(val) + UMC_ADDRCFG_NCOL_BITS_BASE; cs->ucs_nrow_lo = UMC_ADDRCFG_GET_NROW_BITS_LO(val) + UMC_ADDRCFG_NROW_BITS_LO_BASE; cs->ucs_nrow_hi = 0; cs->ucs_nrm = UMC_ADDRCFG_DDR5_GET_NRM_BITS(val); cs->ucs_nbank_groups = UMC_ADDRCFG_GET_NBANKGRP_BITS(val); reg = UMC_ADDRSEL_DDR5(id, regno); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read address select " "register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } cs->ucs_row_hi_bit = 0; cs->ucs_row_low_bit = UMC_ADDRSEL_GET_ROW_LO(val) + UMC_ADDRSEL_ROW_LO_BASE; cs->ucs_bank_bits[4] = UMC_ADDRSEL_GET_BANK4(val) + UMC_ADDRSEL_BANK_BASE; cs->ucs_bank_bits[3] = UMC_ADDRSEL_GET_BANK3(val) + UMC_ADDRSEL_BANK_BASE; cs->ucs_bank_bits[2] = UMC_ADDRSEL_GET_BANK2(val) + UMC_ADDRSEL_BANK_BASE; cs->ucs_bank_bits[1] = UMC_ADDRSEL_GET_BANK1(val) + UMC_ADDRSEL_BANK_BASE; cs->ucs_bank_bits[0] = UMC_ADDRSEL_GET_BANK0(val) + UMC_ADDRSEL_BANK_BASE; reg = UMC_COLSEL_LO_DDR5(id, regno); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read column address " "select low register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } for (uint_t i = 0; i < ZEN_UMC_MAX_COLSEL_PER_REG; i++) { cs->ucs_col_bits[i] = UMC_COLSEL_REMAP_GET_COL(val, i) + UMC_COLSEL_LO_BASE; } reg = UMC_COLSEL_HI_DDR5(id, regno); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read column address " "select high register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } for (uint_t i = 0; i < ZEN_UMC_MAX_COLSEL_PER_REG; i++) { cs->ucs_col_bits[i + ZEN_UMC_MAX_COLSEL_PER_REG] = UMC_COLSEL_REMAP_GET_COL(val, i) + UMC_COLSEL_HI_BASE; } /* * Time for our friend, the RM Selection register. Like in DDR4 we end * up reading everything here, even though most others have reserved * bits here. The intent is that we won't look at the reserved bits * unless something actually points us there. */ reg = UMC_RMSEL_DDR5(id, regno); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read rank multiply " "select register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } /* * DDR5 based devices have a primary and secondary msbs; however, they * only have a single set of rm bits. To normalize things with the DDR4 * subsystem, we copy the primary bits to the secondary so we can use * these the same way in the decoder/encoder. */ cs->ucs_inv_msbs = UMC_RMSEL_DDR5_GET_INV_MSBS(val); cs->ucs_inv_msbs_sec = UMC_RMSEL_DDR5_GET_INV_MSBS_SEC(val); cs->ucs_subchan = UMC_RMSEL_DDR5_GET_SUBCHAN(val) + UMC_RMSEL_DDR5_SUBCHAN_BASE; cs->ucs_rm_bits[3] = UMC_RMSEL_DDR5_GET_RM3(val) + UMC_RMSEL_BASE; cs->ucs_rm_bits[2] = UMC_RMSEL_DDR5_GET_RM2(val) + UMC_RMSEL_BASE; cs->ucs_rm_bits[1] = UMC_RMSEL_DDR5_GET_RM1(val) + UMC_RMSEL_BASE; cs->ucs_rm_bits[0] = UMC_RMSEL_DDR5_GET_RM0(val) + UMC_RMSEL_BASE; bcopy(cs->ucs_rm_bits, cs->ucs_rm_bits_sec, sizeof (cs->ucs_rm_bits)); return (zen_umc_fill_dimm_common(umc, df, chan, dimmno, B_FALSE)); } static void zen_umc_fill_ddr_type(zen_umc_t *umc, zen_umc_chan_t *chan) { umc_dimm_type_t dimm = UMC_DIMM_T_UNKNOWN; uint8_t val; /* * The different UMC styles split into two groups. Those that support * DDR4 and those that support DDR5 (with the hybrid group being in the * DDR5 style camp). While all the values are consistent between * different ones (e.g. reserved values correspond to unsupported * items), we still check types based on the UMC's design type so if we * see something weird, we don't accidentally use an older value. */ val = UMC_UMCCFG_GET_DDR_TYPE(chan->chan_umccfg_raw); switch (umc->umc_fdata->zufd_umc_style) { case ZEN_UMC_UMC_S_DDR4: case ZEN_UMC_UMC_S_DDR4_APU: switch (val) { case UMC_UMCCFG_DDR4_T_DDR4: dimm = UMC_DIMM_T_DDR4; break; case UMC_UMCCFG_DDR4_T_LPDDR4: dimm = UMC_DIMM_T_LPDDR4; break; default: break; } break; case ZEN_UMC_UMC_S_HYBRID_LPDDR5: switch (val) { case UMC_UMCCFG_DDR5_T_LPDDR5: dimm = UMC_DIMM_T_LPDDR5; break; case UMC_UMCCFG_DDR5_T_LPDDR4: dimm = UMC_DIMM_T_LPDDR4; break; default: break; } break; case ZEN_UMC_UMC_S_DDR5: case ZEN_UMC_UMC_S_DDR5_APU: switch (val) { case UMC_UMCCFG_DDR5_T_DDR5: dimm = UMC_DIMM_T_DDR5; break; case UMC_UMCCFG_DDR5_T_LPDDR5: dimm = UMC_DIMM_T_LPDDR5; break; default: break; } break; } chan->chan_type = dimm; } /* * Use the DDR4 frequency table to determine the speed of this. Note that our * hybrid based UMCs use 8 bits for the clock, while the traditional DDR4 ones * only use 7. The caller is responsible for using the right mask for the UMC. */ static void zen_umc_fill_chan_ddr4(zen_umc_chan_t *chan, uint_t mstate, const uint32_t clock) { for (size_t i = 0; i < ARRAY_SIZE(zen_umc_ddr4_map); i++) { if (clock == zen_umc_ddr4_map[i].zufm_reg) { chan->chan_clock[mstate] = zen_umc_ddr4_map[i].zufm_mhz; chan->chan_speed[mstate] = zen_umc_ddr4_map[i].zufm_mts2; break; } } } static void zen_umc_fill_chan_hyb_lpddr5(zen_umc_chan_t *chan, uint_t mstate) { const uint32_t reg = chan->chan_dramcfg_raw[mstate]; const uint32_t wck = UMC_DRAMCFG_HYB_GET_WCLKRATIO(reg); const uint32_t clock = UMC_DRAMCFG_HYB_GET_MEMCLK(reg); boolean_t twox; switch (wck) { case UMC_DRAMCFG_WCLKRATIO_1TO2: twox = B_TRUE; break; case UMC_DRAMCFG_WCLKRATIO_1TO4: twox = B_FALSE; break; default: return; } for (size_t i = 0; i < ARRAY_SIZE(zen_umc_lpddr5_map); i++) { if (clock == zen_umc_lpddr5_map[i].zufm_reg) { chan->chan_clock[mstate] = zen_umc_lpddr5_map[i].zufm_mhz; if (twox) { chan->chan_speed[mstate] = zen_umc_lpddr5_map[i].zufm_mts2; } else { chan->chan_speed[mstate] = zen_umc_lpddr5_map[i].zufm_mts4; } break; } } } /* * Determine the current operating frequency of the channel. This varies based * upon the type of UMC that we're operating on as there are multiple ways to * determine this. There are up to four memory P-states that exist in the UMC. * This grabs it for a single P-state at a time. * * Unlike other things, if we cannot determine the frequency of the clock or * transfer speed, we do not consider this fatal because that does not stop * decoding. It only means that we cannot give a bit of useful information to * topo. */ static void zen_umc_fill_chan_freq(zen_umc_t *umc, zen_umc_chan_t *chan, uint_t mstate) { const uint32_t cfg = chan->chan_dramcfg_raw[mstate]; const umc_dimm_type_t dimm_type = chan->chan_type; switch (umc->umc_fdata->zufd_umc_style) { case ZEN_UMC_UMC_S_HYBRID_LPDDR5: if (dimm_type == UMC_DIMM_T_LPDDR5) { zen_umc_fill_chan_hyb_lpddr5(chan, mstate); } else if (dimm_type != UMC_DIMM_T_LPDDR4) { zen_umc_fill_chan_ddr4(chan, mstate, UMC_DRAMCFG_HYB_GET_MEMCLK(cfg)); } break; case ZEN_UMC_UMC_S_DDR4: case ZEN_UMC_UMC_S_DDR4_APU: zen_umc_fill_chan_ddr4(chan, mstate, UMC_DRAMCFG_DDR4_GET_MEMCLK(cfg)); break; case ZEN_UMC_UMC_S_DDR5: case ZEN_UMC_UMC_S_DDR5_APU: chan->chan_clock[mstate] = UMC_DRAMCFG_DDR5_GET_MEMCLK(cfg); if (dimm_type == UMC_DIMM_T_DDR5) { chan->chan_speed[mstate] = 2 * chan->chan_clock[mstate]; } else if (dimm_type == UMC_DIMM_T_LPDDR5) { switch (UMC_DRAMCFG_LPDDR5_GET_WCKRATIO(cfg)) { case UMC_DRAMCFG_WCLKRATIO_1TO2: chan->chan_speed[mstate] = 2 * chan->chan_clock[mstate]; break; case UMC_DRAMCFG_WCLKRATIO_1TO4: chan->chan_speed[mstate] = 4 * chan->chan_clock[mstate]; break; default: break; } } break; } } /* * Fill common channel information. While the locations of many of the registers * changed between the DDR4-capable and DDR5-capable devices, the actual * contents are the same so we process them together. */ static boolean_t zen_umc_fill_chan_hash(zen_umc_t *umc, zen_umc_df_t *df, zen_umc_chan_t *chan, boolean_t ddr4) { int ret; smn_reg_t reg; uint32_t val; const umc_chan_hash_flags_t flags = umc->umc_fdata->zufd_chan_hash; const uint32_t id = chan->chan_logid; umc_chan_hash_t *chash = &chan->chan_hash; chash->uch_flags = flags; if ((flags & UMC_CHAN_HASH_F_BANK) != 0) { for (uint_t i = 0; i < ZEN_UMC_MAX_CHAN_BANK_HASH; i++) { umc_bank_hash_t *bank = &chash->uch_bank_hashes[i]; if (ddr4) { reg = UMC_BANK_HASH_DDR4(id, i); } else { reg = UMC_BANK_HASH_DDR5(id, i); } if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read " "bank hash register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } bank->ubh_row_xor = UMC_BANK_HASH_GET_ROW(val); bank->ubh_col_xor = UMC_BANK_HASH_GET_COL(val); bank->ubh_en = UMC_BANK_HASH_GET_EN(val); } } if ((flags & UMC_CHAN_HASH_F_RM) != 0) { for (uint_t i = 0; i < ZEN_UMC_MAX_CHAN_RM_HASH; i++) { uint64_t addr; umc_addr_hash_t *rm = &chash->uch_rm_hashes[i]; if (ddr4) { reg = UMC_RANK_HASH_DDR4(id, i); } else { reg = UMC_RANK_HASH_DDR5(id, i); } if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read " "rm hash register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } addr = UMC_RANK_HASH_GET_ADDR(val); rm->uah_addr_xor = addr << UMC_RANK_HASH_SHIFT; rm->uah_en = UMC_RANK_HASH_GET_EN(val); if (ddr4 || (umc->umc_fdata->zufd_flags & ZEN_UMC_FAM_F_UMC_EADDR) == 0) { continue; } reg = UMC_RANK_HASH_EXT_DDR5(id, i); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read " "rm hash ext register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } addr = UMC_RANK_HASH_EXT_GET_ADDR(val); rm->uah_addr_xor |= addr << UMC_RANK_HASH_EXT_ADDR_SHIFT; } } if ((flags & UMC_CHAN_HASH_F_PC) != 0) { umc_pc_hash_t *pc = &chash->uch_pc_hash; if (ddr4) { reg = UMC_PC_HASH_DDR4(id); } else { reg = UMC_PC_HASH_DDR5(id); } if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read pc hash " "register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } pc->uph_row_xor = UMC_PC_HASH_GET_ROW(val); pc->uph_col_xor = UMC_PC_HASH_GET_COL(val); pc->uph_en = UMC_PC_HASH_GET_EN(val); if (ddr4) { reg = UMC_PC_HASH2_DDR4(id); } else { reg = UMC_PC_HASH2_DDR5(id); } if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read pc hash " "2 register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } pc->uph_bank_xor = UMC_PC_HASH2_GET_BANK(val); } if ((flags & UMC_CHAN_HASH_F_CS) != 0) { for (uint_t i = 0; i < ZEN_UMC_MAX_CHAN_CS_HASH; i++) { uint64_t addr; umc_addr_hash_t *rm = &chash->uch_cs_hashes[i]; if (ddr4) { reg = UMC_CS_HASH_DDR4(id, i); } else { reg = UMC_CS_HASH_DDR5(id, i); } if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read " "cs hash register %x", SMN_REG_ADDR(reg)); return (B_FALSE); } addr = UMC_CS_HASH_GET_ADDR(val); rm->uah_addr_xor = addr << UMC_CS_HASH_SHIFT; rm->uah_en = UMC_CS_HASH_GET_EN(val); if (ddr4 || (umc->umc_fdata->zufd_flags & ZEN_UMC_FAM_F_UMC_EADDR) == 0) { continue; } reg = UMC_CS_HASH_EXT_DDR5(id, i); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read " "cs hash ext register %x", SMN_REG_ADDR(reg)); return (B_FALSE); } addr = UMC_CS_HASH_EXT_GET_ADDR(val); rm->uah_addr_xor |= addr << UMC_CS_HASH_EXT_ADDR_SHIFT; } } return (B_TRUE); } /* * This fills in settings that we care about which are valid for the entire * channel and are the same between DDR4/5 capable devices. */ static boolean_t zen_umc_fill_chan(zen_umc_t *umc, zen_umc_df_t *df, zen_umc_chan_t *chan) { uint32_t val; smn_reg_t reg; const uint32_t id = chan->chan_logid; int ret; boolean_t ddr4; if (umc->umc_fdata->zufd_umc_style == ZEN_UMC_UMC_S_DDR4 || umc->umc_fdata->zufd_umc_style == ZEN_UMC_UMC_S_DDR4_APU) { ddr4 = B_TRUE; } else { ddr4 = B_FALSE; } /* * Begin by gathering all of the information related to hashing. What is * valid here varies based on the actual chip family and then the * registers vary based on DDR4 and DDR5. */ if (!zen_umc_fill_chan_hash(umc, df, chan, ddr4)) { return (B_FALSE); } reg = UMC_UMCCFG(id); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read UMC " "configuration register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } chan->chan_umccfg_raw = val; if (UMC_UMCCFG_GET_ECC_EN(val)) { chan->chan_flags |= UMC_CHAN_F_ECC_EN; } /* * Grab the DRAM configuration register. This can be used to determine * the frequency and speed of the memory channel. At this time we only * capture Memory P-state 0. */ reg = UMC_DRAMCFG(id, 0); /* * This register contains information to determine the type of DIMM. * All DIMMs in the channel must be the same type so we leave this * setting on the channel. Once we have that, we proceed to obtain the * currently configuration information for the DRAM in each memory * P-state. */ zen_umc_fill_ddr_type(umc, chan); for (uint_t i = 0; i < ZEN_UMC_NMEM_PSTATES; i++) { chan->chan_clock[i] = ZEN_UMC_UNKNOWN_FREQ; chan->chan_speed[i] = ZEN_UMC_UNKNOWN_FREQ; reg = UMC_DRAMCFG(id, i); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read DRAM " "Configuration register P-state %u %x: %d", i, SMN_REG_ADDR(reg), ret); return (B_FALSE); } chan->chan_dramcfg_raw[i] = val; zen_umc_fill_chan_freq(umc, chan, i); } /* * Grab data that we can use to determine if we're scrambling or * encrypting regions of memory. */ reg = UMC_DATACTL(id); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read data control " "register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } chan->chan_datactl_raw = val; if (UMC_DATACTL_GET_SCRAM_EN(val)) { chan->chan_flags |= UMC_CHAN_F_SCRAMBLE_EN; } if (UMC_DATACTL_GET_ENCR_EN(val)) { chan->chan_flags |= UMC_CHAN_F_ENCR_EN; } /* * At the moment we snapshot the raw ECC control information. When we do * further work of making this a part of the MCA/X decoding, we'll want * to further take this apart for syndrome decoding. Until then, simply * cache it for future us and observability. */ reg = UMC_ECCCTL(id); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read ECC control " "register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } chan->chan_eccctl_raw = val; /* * Read and snapshot the UMC capability registers for debugging in the * future. */ reg = UMC_UMCCAP(id); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read UMC cap" "register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } chan->chan_umccap_raw = val; reg = UMC_UMCCAP_HI(id); if ((ret = amdzen_c_smn_read(df->zud_dfno, reg, &val)) != 0) { dev_err(umc->umc_dip, CE_WARN, "failed to read UMC cap high " "register %x: %d", SMN_REG_ADDR(reg), ret); return (B_FALSE); } chan->chan_umccap_hi_raw = val; return (B_TRUE); } static int zen_umc_fill_umc_cb(const uint_t dfno, const uint32_t fabid, const uint32_t instid, void *arg) { zen_umc_t *umc = arg; zen_umc_df_t *df = &umc->umc_dfs[dfno]; zen_umc_chan_t *chan = &df->zud_chan[df->zud_nchan]; df->zud_nchan++; VERIFY3U(df->zud_nchan, <=, ZEN_UMC_MAX_UMCS); /* * The data fabric is generally organized such that all UMC entries * should be continuous in their fabric ID space; however, we don't * want to rely on specific ID locations. The UMC SMN addresses are * organized in a relative order. To determine the SMN ID to use (the * chan_logid) we end up making the following assumptions: * * o The iteration order will always be from the lowest component ID * to the highest component ID. * o The relative order that we encounter will be the same as the SMN * order. That is, the first thing we find (regardless of component * ID) will be SMN UMC entry 0, the next 1, etc. */ chan->chan_logid = df->zud_nchan - 1; chan->chan_fabid = fabid; chan->chan_instid = instid; chan->chan_nrules = umc->umc_fdata->zufd_cs_nrules; for (uint_t i = 0; i < umc->umc_fdata->zufd_cs_nrules; i++) { if (zen_umc_read_dram_rule(umc, dfno, instid, i, &chan->chan_rules[i]) != 0) { return (-1); } } for (uint_t i = 0; i < umc->umc_fdata->zufd_cs_nrules - 1; i++) { int ret; uint32_t offset; uint64_t t; df_reg_def_t off_reg; chan_offset_t *offp = &chan->chan_offsets[i]; switch (umc->umc_df_rev) { case DF_REV_2: case DF_REV_3: case DF_REV_3P5: ASSERT3U(i, ==, 0); off_reg = DF_DRAM_OFFSET_V2; break; case DF_REV_4: off_reg = DF_DRAM_OFFSET_V4(i); break; default: dev_err(umc->umc_dip, CE_WARN, "!encountered " "unsupported DF revision processing DRAM Offsets: " "0x%x", umc->umc_df_rev); return (-1); } if ((ret = amdzen_c_df_read32(dfno, instid, off_reg, &offset)) != 0) { dev_err(umc->umc_dip, CE_WARN, "!failed to read DRAM " "offset %u on 0x%x/0x%x: %d", i, dfno, instid, ret); return (-1); } offp->cho_raw = offset; offp->cho_valid = DF_DRAM_OFFSET_GET_EN(offset); switch (umc->umc_df_rev) { case DF_REV_2: t = DF_DRAM_OFFSET_V2_GET_OFFSET(offset); break; case DF_REV_3: case DF_REV_3P5: t = DF_DRAM_OFFSET_V3_GET_OFFSET(offset); break; case DF_REV_4: t = DF_DRAM_OFFSET_V4_GET_OFFSET(offset); break; default: dev_err(umc->umc_dip, CE_WARN, "!encountered " "unsupported DF revision processing DRAM Offsets: " "0x%x", umc->umc_df_rev); return (-1); } offp->cho_offset = t << DF_DRAM_OFFSET_SHIFT; } /* * If this platform supports our favorete Zen 3 6-channel hash special * then we need to grab the NP2 configuration registers. This will only * be referenced if this channel is actually being used for a 6-channel * hash, so even if the contents are weird that should still be ok. */ if ((umc->umc_fdata->zufd_flags & ZEN_UMC_FAM_F_NP2) != 0) { uint32_t np2; int ret; if ((ret = amdzen_c_df_read32(dfno, instid, DF_NP2_CONFIG_V3, &np2)) != 0) { dev_err(umc->umc_dip, CE_WARN, "!failed to read NP2 " "config: %d", ret); return (-1); } chan->chan_np2_raw = np2; chan->chan_np2_space0 = DF_NP2_CONFIG_V3_GET_SPACE0(np2); } /* * Now that we have everything we need from the data fabric, read out * the rest of what we need from the UMC channel data in SMN register * space. */ switch (umc->umc_fdata->zufd_umc_style) { case ZEN_UMC_UMC_S_DDR4: case ZEN_UMC_UMC_S_DDR4_APU: for (uint_t i = 0; i < ZEN_UMC_MAX_DIMMS; i++) { if (!zen_umc_fill_chan_dimm_ddr4(umc, df, chan, i)) { return (-1); } } break; case ZEN_UMC_UMC_S_HYBRID_LPDDR5: case ZEN_UMC_UMC_S_DDR5: case ZEN_UMC_UMC_S_DDR5_APU: for (uint_t i = 0; i < ZEN_UMC_MAX_DIMMS; i++) { for (uint_t r = 0; r < ZEN_UMC_MAX_CS_PER_DIMM; r++) { if (!zen_umc_fill_chan_rank_ddr5(umc, df, chan, i, r)) { return (-1); } } } break; default: dev_err(umc->umc_dip, CE_WARN, "!encountered unsupported " "Zen family: 0x%x", umc->umc_fdata->zufd_umc_style); return (-1); } if (!zen_umc_fill_chan(umc, df, chan)) { return (-1); } return (0); } /* * Today there are no privileges for the memory controller information, it is * restricted based on file system permissions. */ static int zen_umc_open(dev_t *devp, int flag, int otyp, cred_t *credp) { zen_umc_t *umc = zen_umc; if ((flag & (FEXCL | FNDELAY | FNONBLOCK | FWRITE)) != 0) { return (EINVAL); } if (otyp != OTYP_CHR) { return (EINVAL); } if (getminor(*devp) >= umc->umc_ndfs) { return (ENXIO); } return (0); } static void zen_umc_ioctl_decode(zen_umc_t *umc, mc_encode_ioc_t *encode) { zen_umc_decoder_t dec; uint32_t sock, die, comp; bzero(&dec, sizeof (dec)); if (!zen_umc_decode_pa(umc, encode->mcei_pa, &dec)) { encode->mcei_err = (uint32_t)dec.dec_fail; encode->mcei_errdata = dec.dec_fail_data; return; } encode->mcei_errdata = 0; encode->mcei_err = 0; encode->mcei_chan_addr = dec.dec_norm_addr; encode->mcei_rank_addr = UINT64_MAX; encode->mcei_board = 0; zen_fabric_id_decompose(&umc->umc_decomp, dec.dec_targ_fabid, &sock, &die, &comp); encode->mcei_chip = sock; encode->mcei_die = die; encode->mcei_mc = dec.dec_umc_chan->chan_logid; encode->mcei_chan = 0; encode->mcei_dimm = dec.dec_dimm_no; encode->mcei_row = dec.dec_dimm_row; encode->mcei_column = dec.dec_dimm_col; /* * We don't have a logical rank that something matches to, we have the * actual chip-select and rank multiplication. If we could figure out * how to transform that into an actual rank, that'd be grand. */ encode->mcei_rank = UINT8_MAX; encode->mcei_cs = dec.dec_dimm_csno; encode->mcei_rm = dec.dec_dimm_rm; encode->mcei_bank = dec.dec_dimm_bank; encode->mcei_bank_group = dec.dec_dimm_bank_group; encode->mcei_subchan = dec.dec_dimm_subchan; } static void umc_decoder_pack(zen_umc_t *umc) { char *buf = NULL; size_t len = 0; ASSERT(MUTEX_HELD(&umc->umc_nvl_lock)); if (umc->umc_decoder_buf != NULL) { return; } if (umc->umc_decoder_nvl == NULL) { umc->umc_decoder_nvl = zen_umc_dump_decoder(umc); if (umc->umc_decoder_nvl == NULL) { return; } } if (nvlist_pack(umc->umc_decoder_nvl, &buf, &len, NV_ENCODE_XDR, KM_NOSLEEP_LAZY) != 0) { return; } umc->umc_decoder_buf = buf; umc->umc_decoder_len = len; } static int zen_umc_ioctl(dev_t dev, int cmd, intptr_t arg, int mode, cred_t *credp, int *rvalp) { int ret; zen_umc_t *umc = zen_umc; mc_encode_ioc_t encode; mc_snapshot_info_t info; if (getminor(dev) >= umc->umc_ndfs) { return (ENXIO); } switch (cmd) { case MC_IOC_DECODE_PA: if (crgetzoneid(credp) != GLOBAL_ZONEID || drv_priv(credp) != 0) { ret = EPERM; break; } if (ddi_copyin((void *)arg, &encode, sizeof (encode), mode & FKIOCTL) != 0) { ret = EFAULT; break; } zen_umc_ioctl_decode(umc, &encode); ret = 0; if (ddi_copyout(&encode, (void *)arg, sizeof (encode), mode & FKIOCTL) != 0) { ret = EFAULT; break; } break; case MC_IOC_DECODE_SNAPSHOT_INFO: mutex_enter(&umc->umc_nvl_lock); umc_decoder_pack(umc); if (umc->umc_decoder_buf == NULL) { mutex_exit(&umc->umc_nvl_lock); ret = EIO; break; } if (umc->umc_decoder_len > UINT32_MAX) { mutex_exit(&umc->umc_nvl_lock); ret = EOVERFLOW; break; } info.mcs_size = umc->umc_decoder_len; info.mcs_gen = 0; if (ddi_copyout(&info, (void *)arg, sizeof (info), mode & FKIOCTL) != 0) { mutex_exit(&umc->umc_nvl_lock); ret = EFAULT; break; } mutex_exit(&umc->umc_nvl_lock); ret = 0; break; case MC_IOC_DECODE_SNAPSHOT: mutex_enter(&umc->umc_nvl_lock); umc_decoder_pack(umc); if (umc->umc_decoder_buf == NULL) { mutex_exit(&umc->umc_nvl_lock); ret = EIO; break; } if (ddi_copyout(umc->umc_decoder_buf, (void *)arg, umc->umc_decoder_len, mode & FKIOCTL) != 0) { mutex_exit(&umc->umc_nvl_lock); ret = EFAULT; break; } mutex_exit(&umc->umc_nvl_lock); ret = 0; break; default: ret = ENOTTY; break; } return (ret); } static int zen_umc_close(dev_t dev, int flag, int otyp, cred_t *credp) { return (0); } static void zen_umc_cleanup(zen_umc_t *umc) { nvlist_free(umc->umc_decoder_nvl); umc->umc_decoder_nvl = NULL; if (umc->umc_decoder_buf != NULL) { kmem_free(umc->umc_decoder_buf, umc->umc_decoder_len); umc->umc_decoder_buf = NULL; umc->umc_decoder_len = 0; } if (umc->umc_dip != NULL) { ddi_remove_minor_node(umc->umc_dip, NULL); } mutex_destroy(&umc->umc_nvl_lock); kmem_free(umc, sizeof (zen_umc_t)); } static int zen_umc_attach(dev_info_t *dip, ddi_attach_cmd_t cmd) { int ret; zen_umc_t *umc; if (cmd == DDI_RESUME) { return (DDI_SUCCESS); } else if (cmd != DDI_ATTACH) { return (DDI_FAILURE); } if (zen_umc != NULL) { dev_err(dip, CE_WARN, "!zen_umc is already attached to a " "dev_info_t: %p", zen_umc->umc_dip); return (DDI_FAILURE); } /* * To get us going, we need to do several bits of set up. First, we need * to use the knowledge about the actual hardware that we're using to * encode a bunch of different data: * * o The set of register styles and extra hardware features that exist * on the hardware platform. * o The number of actual rules there are for the CCMs and UMCs. * o How many actual things exist (DFs, etc.) * o Useful fabric and instance IDs for all of the different UMC * entries so we can actually talk to them. * * Only once we have all the above will we go dig into the actual data. */ umc = kmem_zalloc(sizeof (zen_umc_t), KM_SLEEP); mutex_init(&umc->umc_nvl_lock, NULL, MUTEX_DRIVER, NULL); umc->umc_family = chiprev_family(cpuid_getchiprev(CPU)); umc->umc_ndfs = amdzen_c_df_count(); umc->umc_dip = dip; if (!zen_umc_identify(umc)) { dev_err(dip, CE_WARN, "!encountered unsupported CPU"); goto err; } umc->umc_df_rev = amdzen_c_df_rev(); switch (umc->umc_df_rev) { case DF_REV_2: case DF_REV_3: case DF_REV_3P5: case DF_REV_4: break; default: dev_err(dip, CE_WARN, "!encountered unknown DF revision: %x", umc->umc_df_rev); goto err; } if ((ret = amdzen_c_df_fabric_decomp(&umc->umc_decomp)) != 0) { dev_err(dip, CE_WARN, "!failed to get fabric decomposition: %d", ret); } umc->umc_tom = rdmsr(MSR_AMD_TOM); umc->umc_tom2 = rdmsr(MSR_AMD_TOM2); /* * For each DF, start by reading all of the data that we need from it. * This involves finding a target CCM, reading all of the rules, * ancillary settings, and related. Then we'll do a pass over all of the * actual UMC targets there. */ for (uint_t i = 0; i < umc->umc_ndfs; i++) { if (amdzen_c_df_iter(i, ZEN_DF_TYPE_CCM_CPU, zen_umc_fill_ccm_cb, umc) < 0 || amdzen_c_df_iter(i, ZEN_DF_TYPE_CS_UMC, zen_umc_fill_umc_cb, umc) != 0) { goto err; } } /* * Create a minor node for each df that we encounter. */ for (uint_t i = 0; i < umc->umc_ndfs; i++) { int ret; char minor[64]; (void) snprintf(minor, sizeof (minor), "mc-umc-%u", i); if ((ret = ddi_create_minor_node(umc->umc_dip, minor, S_IFCHR, i, "ddi_mem_ctrl", 0)) != 0) { dev_err(dip, CE_WARN, "!failed to create minor %s: %d", minor, ret); goto err; } } zen_umc = umc; return (DDI_SUCCESS); err: zen_umc_cleanup(umc); return (DDI_FAILURE); } static int zen_umc_getinfo(dev_info_t *dip, ddi_info_cmd_t cmd, void *arg, void **resultp) { zen_umc_t *umc; if (zen_umc == NULL || zen_umc->umc_dip == NULL) { return (DDI_FAILURE); } umc = zen_umc; switch (cmd) { case DDI_INFO_DEVT2DEVINFO: *resultp = (void *)umc->umc_dip; break; case DDI_INFO_DEVT2INSTANCE: *resultp = (void *)(uintptr_t)ddi_get_instance( umc->umc_dip); break; default: return (DDI_FAILURE); } return (DDI_SUCCESS); } static int zen_umc_detach(dev_info_t *dip, ddi_detach_cmd_t cmd) { zen_umc_t *umc; if (cmd == DDI_SUSPEND) { return (DDI_SUCCESS); } else if (cmd != DDI_DETACH) { return (DDI_FAILURE); } if (zen_umc == NULL) { dev_err(dip, CE_WARN, "!asked to detach zen_umc, but it " "was never successfully attached"); return (DDI_FAILURE); } umc = zen_umc; zen_umc = NULL; zen_umc_cleanup(umc); return (DDI_SUCCESS); } static struct cb_ops zen_umc_cb_ops = { .cb_open = zen_umc_open, .cb_close = zen_umc_close, .cb_strategy = nodev, .cb_print = nodev, .cb_dump = nodev, .cb_read = nodev, .cb_write = nodev, .cb_ioctl = zen_umc_ioctl, .cb_devmap = nodev, .cb_mmap = nodev, .cb_segmap = nodev, .cb_chpoll = nochpoll, .cb_prop_op = ddi_prop_op, .cb_flag = D_MP, .cb_rev = CB_REV, .cb_aread = nodev, .cb_awrite = nodev }; static struct dev_ops zen_umc_dev_ops = { .devo_rev = DEVO_REV, .devo_refcnt = 0, .devo_getinfo = zen_umc_getinfo, .devo_identify = nulldev, .devo_probe = nulldev, .devo_attach = zen_umc_attach, .devo_detach = zen_umc_detach, .devo_reset = nodev, .devo_quiesce = ddi_quiesce_not_needed, .devo_cb_ops = &zen_umc_cb_ops }; static struct modldrv zen_umc_modldrv = { .drv_modops = &mod_driverops, .drv_linkinfo = "AMD Zen Unified Memory Controller", .drv_dev_ops = &zen_umc_dev_ops }; static struct modlinkage zen_umc_modlinkage = { .ml_rev = MODREV_1, .ml_linkage = { &zen_umc_modldrv, NULL } }; int _init(void) { return (mod_install(&zen_umc_modlinkage)); } int _info(struct modinfo *modinfop) { return (mod_info(&zen_umc_modlinkage, modinfop)); } int _fini(void) { return (mod_remove(&zen_umc_modlinkage)); }