1Heterogeneous Memory Management (HMM) 2 3Transparently allow any component of a program to use any memory region of said 4program with a device without using device specific memory allocator. This is 5becoming a requirement to simplify the use of advance heterogeneous computing 6where GPU, DSP or FPGA are use to perform various computations. 7 8This document is divided as follow, in the first section i expose the problems 9related to the use of a device specific allocator. The second section i expose 10the hardware limitations that are inherent to many platforms. The third section 11gives an overview of HMM designs. The fourth section explains how CPU page- 12table mirroring works and what is HMM purpose in this context. Fifth section 13deals with how device memory is represented inside the kernel. Finaly the last 14section present the new migration helper that allow to leverage the device DMA 15engine. 16 17 181) Problems of using device specific memory allocator: 192) System bus, device memory characteristics 203) Share address space and migration 214) Address space mirroring implementation and API 225) Represent and manage device memory from core kernel point of view 236) Migrate to and from device memory 247) Memory cgroup (memcg) and rss accounting 25 26 27------------------------------------------------------------------------------- 28 291) Problems of using device specific memory allocator: 30 31Device with large amount of on board memory (several giga bytes) like GPU have 32historically manage their memory through dedicated driver specific API. This 33creates a disconnect between memory allocated and managed by device driver and 34regular application memory (private anonymous, share memory or regular file 35back memory). From here on i will refer to this aspect as split address space. 36I use share address space to refer to the opposite situation ie one in which 37any memory region can be use by device transparently. 38 39Split address space because device can only access memory allocated through the 40device specific API. This imply that all memory object in a program are not 41equal from device point of view which complicate large program that rely on a 42wide set of libraries. 43 44Concretly this means that code that wants to leverage device like GPU need to 45copy object between genericly allocated memory (malloc, mmap private/share/) 46and memory allocated through the device driver API (this still end up with an 47mmap but of the device file). 48 49For flat dataset (array, grid, image, ...) this isn't too hard to achieve but 50complex data-set (list, tree, ...) are hard to get right. Duplicating a complex 51data-set need to re-map all the pointer relations between each of its elements. 52This is error prone and program gets harder to debug because of the duplicate 53data-set. 54 55Split address space also means that library can not transparently use data they 56are getting from core program or other library and thus each library might have 57to duplicate its input data-set using specific memory allocator. Large project 58suffer from this and waste resources because of the various memory copy. 59 60Duplicating each library API to accept as input or output memory allocted by 61each device specific allocator is not a viable option. It would lead to a 62combinatorial explosions in the library entry points. 63 64Finaly with the advance of high level language constructs (in C++ but in other 65language too) it is now possible for compiler to leverage GPU or other devices 66without even the programmer knowledge. Some of compiler identified patterns are 67only do-able with a share address. It is as well more reasonable to use a share 68address space for all the other patterns. 69 70 71------------------------------------------------------------------------------- 72 732) System bus, device memory characteristics 74 75System bus cripple share address due to few limitations. Most system bus only 76allow basic memory access from device to main memory, even cache coherency is 77often optional. Access to device memory from CPU is even more limited, most 78often than not it is not cache coherent. 79 80If we only consider the PCIE bus than device can access main memory (often 81through an IOMMU) and be cache coherent with the CPUs. However it only allows 82a limited set of atomic operation from device on main memory. This is worse 83in the other direction the CPUs can only access a limited range of the device 84memory and can not perform atomic operations on it. Thus device memory can not 85be consider like regular memory from kernel point of view. 86 87Another crippling factor is the limited bandwidth (~32GBytes/s with PCIE 4.0 88and 16 lanes). This is 33 times less that fastest GPU memory (1 TBytes/s). 89The final limitation is latency, access to main memory from the device has an 90order of magnitude higher latency than when the device access its own memory. 91 92Some platform are developing new system bus or additions/modifications to PCIE 93to address some of those limitations (OpenCAPI, CCIX). They mainly allow two 94way cache coherency between CPU and device and allow all atomic operations the 95architecture supports. Saddly not all platform are following this trends and 96some major architecture are left without hardware solutions to those problems. 97 98So for share address space to make sense not only we must allow device to 99access any memory memory but we must also permit any memory to be migrated to 100device memory while device is using it (blocking CPU access while it happens). 101 102 103------------------------------------------------------------------------------- 104 1053) Share address space and migration 106 107HMM intends to provide two main features. First one is to share the address 108space by duplication the CPU page table into the device page table so same 109address point to same memory and this for any valid main memory address in 110the process address space. 111 112To achieve this, HMM offer a set of helpers to populate the device page table 113while keeping track of CPU page table updates. Device page table updates are 114not as easy as CPU page table updates. To update the device page table you must 115allow a buffer (or use a pool of pre-allocated buffer) and write GPU specifics 116commands in it to perform the update (unmap, cache invalidations and flush, 117...). This can not be done through common code for all device. Hence why HMM 118provides helpers to factor out everything that can be while leaving the gory 119details to the device driver. 120 121The second mechanism HMM provide is a new kind of ZONE_DEVICE memory that does 122allow to allocate a struct page for each page of the device memory. Those page 123are special because the CPU can not map them. They however allow to migrate 124main memory to device memory using exhisting migration mechanism and everything 125looks like if page was swap out to disk from CPU point of view. Using a struct 126page gives the easiest and cleanest integration with existing mm mechanisms. 127Again here HMM only provide helpers, first to hotplug new ZONE_DEVICE memory 128for the device memory and second to perform migration. Policy decision of what 129and when to migrate things is left to the device driver. 130 131Note that any CPU access to a device page trigger a page fault and a migration 132back to main memory ie when a page backing an given address A is migrated from 133a main memory page to a device page then any CPU access to address A trigger a 134page fault and initiate a migration back to main memory. 135 136 137With this two features, HMM not only allow a device to mirror a process address 138space and keeps both CPU and device page table synchronize, but also allow to 139leverage device memory by migrating part of data-set that is actively use by a 140device. 141 142 143------------------------------------------------------------------------------- 144 1454) Address space mirroring implementation and API 146 147Address space mirroring main objective is to allow to duplicate range of CPU 148page table into a device page table and HMM helps keeping both synchronize. A 149device driver that want to mirror a process address space must start with the 150registration of an hmm_mirror struct: 151 152 int hmm_mirror_register(struct hmm_mirror *mirror, 153 struct mm_struct *mm); 154 int hmm_mirror_register_locked(struct hmm_mirror *mirror, 155 struct mm_struct *mm); 156 157The locked variant is to be use when the driver is already holding the mmap_sem 158of the mm in write mode. The mirror struct has a set of callback that are use 159to propagate CPU page table: 160 161 struct hmm_mirror_ops { 162 /* sync_cpu_device_pagetables() - synchronize page tables 163 * 164 * @mirror: pointer to struct hmm_mirror 165 * @update_type: type of update that occurred to the CPU page table 166 * @start: virtual start address of the range to update 167 * @end: virtual end address of the range to update 168 * 169 * This callback ultimately originates from mmu_notifiers when the CPU 170 * page table is updated. The device driver must update its page table 171 * in response to this callback. The update argument tells what action 172 * to perform. 173 * 174 * The device driver must not return from this callback until the device 175 * page tables are completely updated (TLBs flushed, etc); this is a 176 * synchronous call. 177 */ 178 void (*update)(struct hmm_mirror *mirror, 179 enum hmm_update action, 180 unsigned long start, 181 unsigned long end); 182 }; 183 184Device driver must perform update to the range following action (turn range 185read only, or fully unmap, ...). Once driver callback returns the device must 186be done with the update. 187 188 189When device driver wants to populate a range of virtual address it can use 190either: 191 int hmm_vma_get_pfns(struct vm_area_struct *vma, 192 struct hmm_range *range, 193 unsigned long start, 194 unsigned long end, 195 hmm_pfn_t *pfns); 196 int hmm_vma_fault(struct vm_area_struct *vma, 197 struct hmm_range *range, 198 unsigned long start, 199 unsigned long end, 200 hmm_pfn_t *pfns, 201 bool write, 202 bool block); 203 204First one (hmm_vma_get_pfns()) will only fetch present CPU page table entry and 205will not trigger a page fault on missing or non present entry. The second one 206do trigger page fault on missing or read only entry if write parameter is true. 207Page fault use the generic mm page fault code path just like a CPU page fault. 208 209Both function copy CPU page table into their pfns array argument. Each entry in 210that array correspond to an address in the virtual range. HMM provide a set of 211flags to help driver identify special CPU page table entries. 212 213Locking with the update() callback is the most important aspect the driver must 214respect in order to keep things properly synchronize. The usage pattern is : 215 216 int driver_populate_range(...) 217 { 218 struct hmm_range range; 219 ... 220 again: 221 ret = hmm_vma_get_pfns(vma, &range, start, end, pfns); 222 if (ret) 223 return ret; 224 take_lock(driver->update); 225 if (!hmm_vma_range_done(vma, &range)) { 226 release_lock(driver->update); 227 goto again; 228 } 229 230 // Use pfns array content to update device page table 231 232 release_lock(driver->update); 233 return 0; 234 } 235 236The driver->update lock is the same lock that driver takes inside its update() 237callback. That lock must be call before hmm_vma_range_done() to avoid any race 238with a concurrent CPU page table update. 239 240HMM implements all this on top of the mmu_notifier API because we wanted to a 241simpler API and also to be able to perform optimization latter own like doing 242concurrent device update in multi-devices scenario. 243 244HMM also serve as an impedence missmatch between how CPU page table update are 245done (by CPU write to the page table and TLB flushes) from how device update 246their own page table. Device update is a multi-step process, first appropriate 247commands are write to a buffer, then this buffer is schedule for execution on 248the device. It is only once the device has executed commands in the buffer that 249the update is done. Creating and scheduling update command buffer can happen 250concurrently for multiple devices. Waiting for each device to report commands 251as executed is serialize (there is no point in doing this concurrently). 252 253 254------------------------------------------------------------------------------- 255 2565) Represent and manage device memory from core kernel point of view 257 258Several differents design were try to support device memory. First one use 259device specific data structure to keep information about migrated memory and 260HMM hooked itself in various place of mm code to handle any access to address 261that were back by device memory. It turns out that this ended up replicating 262most of the fields of struct page and also needed many kernel code path to be 263updated to understand this new kind of memory. 264 265Thing is most kernel code path never try to access the memory behind a page 266but only care about struct page contents. Because of this HMM switchted to 267directly using struct page for device memory which left most kernel code path 268un-aware of the difference. We only need to make sure that no one ever try to 269map those page from the CPU side. 270 271HMM provide a set of helpers to register and hotplug device memory as a new 272region needing struct page. This is offer through a very simple API: 273 274 struct hmm_devmem *hmm_devmem_add(const struct hmm_devmem_ops *ops, 275 struct device *device, 276 unsigned long size); 277 void hmm_devmem_remove(struct hmm_devmem *devmem); 278 279The hmm_devmem_ops is where most of the important things are: 280 281 struct hmm_devmem_ops { 282 void (*free)(struct hmm_devmem *devmem, struct page *page); 283 int (*fault)(struct hmm_devmem *devmem, 284 struct vm_area_struct *vma, 285 unsigned long addr, 286 struct page *page, 287 unsigned flags, 288 pmd_t *pmdp); 289 }; 290 291The first callback (free()) happens when the last reference on a device page is 292drop. This means the device page is now free and no longer use by anyone. The 293second callback happens whenever CPU try to access a device page which it can 294not do. This second callback must trigger a migration back to system memory. 295 296 297------------------------------------------------------------------------------- 298 2996) Migrate to and from device memory 300 301Because CPU can not access device memory, migration must use device DMA engine 302to perform copy from and to device memory. For this we need a new migration 303helper: 304 305 int migrate_vma(const struct migrate_vma_ops *ops, 306 struct vm_area_struct *vma, 307 unsigned long mentries, 308 unsigned long start, 309 unsigned long end, 310 unsigned long *src, 311 unsigned long *dst, 312 void *private); 313 314Unlike other migration function it works on a range of virtual address, there 315is two reasons for that. First device DMA copy has a high setup overhead cost 316and thus batching multiple pages is needed as otherwise the migration overhead 317make the whole excersie pointless. The second reason is because driver trigger 318such migration base on range of address the device is actively accessing. 319 320The migrate_vma_ops struct define two callbacks. First one (alloc_and_copy()) 321control destination memory allocation and copy operation. Second one is there 322to allow device driver to perform cleanup operation after migration. 323 324 struct migrate_vma_ops { 325 void (*alloc_and_copy)(struct vm_area_struct *vma, 326 const unsigned long *src, 327 unsigned long *dst, 328 unsigned long start, 329 unsigned long end, 330 void *private); 331 void (*finalize_and_map)(struct vm_area_struct *vma, 332 const unsigned long *src, 333 const unsigned long *dst, 334 unsigned long start, 335 unsigned long end, 336 void *private); 337 }; 338 339It is important to stress that this migration helpers allow for hole in the 340virtual address range. Some pages in the range might not be migrated for all 341the usual reasons (page is pin, page is lock, ...). This helper does not fail 342but just skip over those pages. 343 344The alloc_and_copy() might as well decide to not migrate all pages in the 345range (for reasons under the callback control). For those the callback just 346have to leave the corresponding dst entry empty. 347 348Finaly the migration of the struct page might fails (for file back page) for 349various reasons (failure to freeze reference, or update page cache, ...). If 350that happens then the finalize_and_map() can catch any pages that was not 351migrated. Note those page were still copied to new page and thus we wasted 352bandwidth but this is considered as a rare event and a price that we are 353willing to pay to keep all the code simpler. 354 355 356------------------------------------------------------------------------------- 357 3587) Memory cgroup (memcg) and rss accounting 359 360For now device memory is accounted as any regular page in rss counters (either 361anonymous if device page is use for anonymous, file if device page is use for 362file back page or shmem if device page is use for share memory). This is a 363deliberate choice to keep existing application that might start using device 364memory without knowing about it to keep runing unimpacted. 365 366Drawbacks is that OOM killer might kill an application using a lot of device 367memory and not a lot of regular system memory and thus not freeing much system 368memory. We want to gather more real world experience on how application and 369system react under memory pressure in the presence of device memory before 370deciding to account device memory differently. 371 372 373Same decision was made for memory cgroup. Device memory page are accounted 374against same memory cgroup a regular page would be accounted to. This does 375simplify migration to and from device memory. This also means that migration 376back from device memory to regular memory can not fail because it would 377go above memory cgroup limit. We might revisit this choice latter on once we 378get more experience in how device memory is use and its impact on memory 379resource control. 380 381 382Note that device memory can never be pin nor by device driver nor through GUP 383and thus such memory is always free upon process exit. Or when last reference 384is drop in case of share memory or file back memory. 385