1    Notes on the Generic Block Layer Rewrite in Linux 2.5
2    =====================================================
4Notes Written on Jan 15, 2002:
5    Jens Axboe <>
6    Suparna Bhattacharya <>
8Last Updated May 2, 2002
9September 2003: Updated I/O Scheduler portions
10    Nick Piggin <>
14These are some notes describing some aspects of the 2.5 block layer in the
15context of the bio rewrite. The idea is to bring out some of the key
16changes and a glimpse of the rationale behind those changes.
18Please mail corrections & suggestions to
232.5 bio rewrite:
24    Jens Axboe <>
26Many aspects of the generic block layer redesign were driven by and evolved
27over discussions, prior patches and the collective experience of several
28people. See sections 8 and 9 for a list of some related references.
30The following people helped with review comments and inputs for this
32    Christoph Hellwig <>
33    Arjan van de Ven <>
34    Randy Dunlap <>
35    Andre Hedrick <>
37The following people helped with fixes/contributions to the bio patches
38while it was still work-in-progress:
39    David S. Miller <>
42Description of Contents:
451. Scope for tuning of logic to various needs
46  1.1 Tuning based on device or low level driver capabilities
47    - Per-queue parameters
48    - Highmem I/O support
49    - I/O scheduler modularization
50  1.2 Tuning based on high level requirements/capabilities
51    1.2.1 I/O Barriers
52    1.2.2 Request Priority/Latency
53  1.3 Direct access/bypass to lower layers for diagnostics and special
54      device operations
55    1.3.1 Pre-built commands
562. New flexible and generic but minimalist i/o structure or descriptor
57   (instead of using buffer heads at the i/o layer)
58  2.1 Requirements/Goals addressed
59  2.2 The bio struct in detail (multi-page io unit)
60  2.3 Changes in the request structure
613. Using bios
62  3.1 Setup/teardown (allocation, splitting)
63  3.2 Generic bio helper routines
64    3.2.1 Traversing segments and completion units in a request
65    3.2.2 Setting up DMA scatterlists
66    3.2.3 I/O completion
67    3.2.4 Implications for drivers that do not interpret bios (don't handle
68       multiple segments)
69    3.2.5 Request command tagging
70  3.3 I/O submission
714. The I/O scheduler
725. Scalability related changes
73  5.1 Granular locking: Removal of io_request_lock
74  5.2 Prepare for transition to 64 bit sector_t
756. Other Changes/Implications
76  6.1 Partition re-mapping handled by the generic block layer
777. A few tips on migration of older drivers
788. A list of prior/related/impacted patches/ideas
799. Other References/Discussion Threads
83Bio Notes
86Let us discuss the changes in the context of how some overall goals for the
87block layer are addressed.
891. Scope for tuning the generic logic to satisfy various requirements
91The block layer design supports adaptable abstractions to handle common
92processing with the ability to tune the logic to an appropriate extent
93depending on the nature of the device and the requirements of the caller.
94One of the objectives of the rewrite was to increase the degree of tunability
95and to enable higher level code to utilize underlying device/driver
96capabilities to the maximum extent for better i/o performance. This is
97important especially in the light of ever improving hardware capabilities
98and application/middleware software designed to take advantage of these
1011.1 Tuning based on low level device / driver capabilities
103Sophisticated devices with large built-in caches, intelligent i/o scheduling
104optimizations, high memory DMA support, etc may find some of the
105generic processing an overhead, while for less capable devices the
106generic functionality is essential for performance or correctness reasons.
107Knowledge of some of the capabilities or parameters of the device should be
108used at the generic block layer to take the right decisions on
109behalf of the driver.
111How is this achieved ?
113Tuning at a per-queue level:
115i. Per-queue limits/values exported to the generic layer by the driver
117Various parameters that the generic i/o scheduler logic uses are set at
118a per-queue level (e.g maximum request size, maximum number of segments in
119a scatter-gather list, hardsect size)
121Some parameters that were earlier available as global arrays indexed by
122major/minor are now directly associated with the queue. Some of these may
123move into the block device structure in the future. Some characteristics
124have been incorporated into a queue flags field rather than separate fields
125in themselves. There are blk_queue_xxx functions to set the parameters,
126rather than update the fields directly
128Some new queue property settings:
130    blk_queue_bounce_limit(q, u64 dma_address)
131        Enable I/O to highmem pages, dma_address being the
132        limit. No highmem default.
134    blk_queue_max_sectors(q, max_sectors)
135        Sets two variables that limit the size of the request.
137        - The request queue's max_sectors, which is a soft size in
138        units of 512 byte sectors, and could be dynamically varied
139        by the core kernel.
141        - The request queue's max_hw_sectors, which is a hard limit
142        and reflects the maximum size request a driver can handle
143        in units of 512 byte sectors.
145        The default for both max_sectors and max_hw_sectors is
146        255. The upper limit of max_sectors is 1024.
148    blk_queue_max_phys_segments(q, max_segments)
149        Maximum physical segments you can handle in a request. 128
150        default (driver limit). (See 3.2.2)
152    blk_queue_max_hw_segments(q, max_segments)
153        Maximum dma segments the hardware can handle in a request. 128
154        default (host adapter limit, after dma remapping).
155        (See 3.2.2)
157    blk_queue_max_segment_size(q, max_seg_size)
158        Maximum size of a clustered segment, 64kB default.
160    blk_queue_hardsect_size(q, hardsect_size)
161        Lowest possible sector size that the hardware can operate
162        on, 512 bytes default.
164New queue flags:
166    QUEUE_FLAG_CLUSTER (see 3.2.2)
167    QUEUE_FLAG_QUEUED (see 3.2.4)
170ii. High-mem i/o capabilities are now considered the default
172The generic bounce buffer logic, present in 2.4, where the block layer would
173by default copyin/out i/o requests on high-memory buffers to low-memory buffers
174assuming that the driver wouldn't be able to handle it directly, has been
175changed in 2.5. The bounce logic is now applied only for memory ranges
176for which the device cannot handle i/o. A driver can specify this by
177setting the queue bounce limit for the request queue for the device
178(blk_queue_bounce_limit()). This avoids the inefficiencies of the copyin/out
179where a device is capable of handling high memory i/o.
181In order to enable high-memory i/o where the device is capable of supporting
182it, the pci dma mapping routines and associated data structures have now been
183modified to accomplish a direct page -> bus translation, without requiring
184a virtual address mapping (unlike the earlier scheme of virtual address
185-> bus translation). So this works uniformly for high-memory pages (which
186do not have a corresponding kernel virtual address space mapping) and
187low-memory pages.
189Note: Please refer to Documentation/PCI/PCI-DMA-mapping.txt for a discussion
190on PCI high mem DMA aspects and mapping of scatter gather lists, and support
191for 64 bit PCI.
193Special handling is required only for cases where i/o needs to happen on
194pages at physical memory addresses beyond what the device can support. In these
195cases, a bounce bio representing a buffer from the supported memory range
196is used for performing the i/o with copyin/copyout as needed depending on
197the type of the operation. For example, in case of a read operation, the
198data read has to be copied to the original buffer on i/o completion, so a
199callback routine is set up to do this, while for write, the data is copied
200from the original buffer to the bounce buffer prior to issuing the
201operation. Since an original buffer may be in a high memory area that's not
202mapped in kernel virtual addr, a kmap operation may be required for
203performing the copy, and special care may be needed in the completion path
204as it may not be in irq context. Special care is also required (by way of
205GFP flags) when allocating bounce buffers, to avoid certain highmem
206deadlock possibilities.
208It is also possible that a bounce buffer may be allocated from high-memory
209area that's not mapped in kernel virtual addr, but within the range that the
210device can use directly; so the bounce page may need to be kmapped during
211copy operations. [Note: This does not hold in the current implementation,
214There are some situations when pages from high memory may need to
215be kmapped, even if bounce buffers are not necessary. For example a device
216may need to abort DMA operations and revert to PIO for the transfer, in
217which case a virtual mapping of the page is required. For SCSI it is also
218done in some scenarios where the low level driver cannot be trusted to
219handle a single sg entry correctly. The driver is expected to perform the
220kmaps as needed on such occasions using the __bio_kmap_atomic and bio_kmap_irq
221routines as appropriate. A driver could also use the blk_queue_bounce()
222routine on its own to bounce highmem i/o to low memory for specific requests
223if so desired.
225iii. The i/o scheduler algorithm itself can be replaced/set as appropriate
227As in 2.4, it is possible to plugin a brand new i/o scheduler for a particular
228queue or pick from (copy) existing generic schedulers and replace/override
229certain portions of it. The 2.5 rewrite provides improved modularization
230of the i/o scheduler. There are more pluggable callbacks, e.g for init,
231add request, extract request, which makes it possible to abstract specific
232i/o scheduling algorithm aspects and details outside of the generic loop.
233It also makes it possible to completely hide the implementation details of
234the i/o scheduler from block drivers.
236I/O scheduler wrappers are to be used instead of accessing the queue directly.
237See section 4. The I/O scheduler for details.
2391.2 Tuning Based on High level code capabilities
241i. Application capabilities for raw i/o
243This comes from some of the high-performance database/middleware
244requirements where an application prefers to make its own i/o scheduling
245decisions based on an understanding of the access patterns and i/o
248ii. High performance filesystems or other higher level kernel code's
251Kernel components like filesystems could also take their own i/o scheduling
252decisions for optimizing performance. Journalling filesystems may need
253some control over i/o ordering.
255What kind of support exists at the generic block layer for this ?
257The flags and rw fields in the bio structure can be used for some tuning
258from above e.g indicating that an i/o is just a readahead request, or for
259marking barrier requests (discussed next), or priority settings (currently
260unused). As far as user applications are concerned they would need an
261additional mechanism either via open flags or ioctls, or some other upper
262level mechanism to communicate such settings to block.
2641.2.1 I/O Barriers
266There is a way to enforce strict ordering for i/os through barriers.
267All requests before a barrier point must be serviced before the barrier
268request and any other requests arriving after the barrier will not be
269serviced until after the barrier has completed. This is useful for higher
270level control on write ordering, e.g flushing a log of committed updates
271to disk before the corresponding updates themselves.
273A flag in the bio structure, BIO_BARRIER is used to identify a barrier i/o.
274The generic i/o scheduler would make sure that it places the barrier request and
275all other requests coming after it after all the previous requests in the
276queue. Barriers may be implemented in different ways depending on the
277driver. For more details regarding I/O barriers, please read barrier.txt
278in this directory.
2801.2.2 Request Priority/Latency
282Todo/Under discussion:
283Arjan's proposed request priority scheme allows higher levels some broad
284  control (high/med/low) over the priority of an i/o request vs other pending
285  requests in the queue. For example it allows reads for bringing in an
286  executable page on demand to be given a higher priority over pending write
287  requests which haven't aged too much on the queue. Potentially this priority
288  could even be exposed to applications in some manner, providing higher level
289  tunability. Time based aging avoids starvation of lower priority
290  requests. Some bits in the bi_rw flags field in the bio structure are
291  intended to be used for this priority information.
2941.3 Direct Access to Low level Device/Driver Capabilities (Bypass mode)
295    (e.g Diagnostics, Systems Management)
297There are situations where high-level code needs to have direct access to
298the low level device capabilities or requires the ability to issue commands
299to the device bypassing some of the intermediate i/o layers.
300These could, for example, be special control commands issued through ioctl
301interfaces, or could be raw read/write commands that stress the drive's
302capabilities for certain kinds of fitness tests. Having direct interfaces at
303multiple levels without having to pass through upper layers makes
304it possible to perform bottom up validation of the i/o path, layer by
305layer, starting from the media.
307The normal i/o submission interfaces, e.g submit_bio, could be bypassed
308for specially crafted requests which such ioctl or diagnostics
309interfaces would typically use, and the elevator add_request routine
310can instead be used to directly insert such requests in the queue or preferably
311the blk_do_rq routine can be used to place the request on the queue and
312wait for completion. Alternatively, sometimes the caller might just
313invoke a lower level driver specific interface with the request as a
316If the request is a means for passing on special information associated with
317the command, then such information is associated with the request->special
318field (rather than misuse the request->buffer field which is meant for the
319request data buffer's virtual mapping).
321For passing request data, the caller must build up a bio descriptor
322representing the concerned memory buffer if the underlying driver interprets
323bio segments or uses the block layer end*request* functions for i/o
324completion. Alternatively one could directly use the request->buffer field to
325specify the virtual address of the buffer, if the driver expects buffer
326addresses passed in this way and ignores bio entries for the request type
327involved. In the latter case, the driver would modify and manage the
328request->buffer, request->sector and request->nr_sectors or
329request->current_nr_sectors fields itself rather than using the block layer
330end_request or end_that_request_first completion interfaces.
331(See 2.3 or Documentation/block/request.txt for a brief explanation of
332the request structure fields)
334[TBD: end_that_request_last should be usable even in this case;
335Perhaps an end_that_direct_request_first routine could be implemented to make
336handling direct requests easier for such drivers; Also for drivers that
337expect bios, a helper function could be provided for setting up a bio
338corresponding to a data buffer]
340<JENS: I dont understand the above, why is end_that_request_first() not
341usable? Or _last for that matter. I must be missing something>
342<SUP: What I meant here was that if the request doesn't have a bio, then
343 end_that_request_first doesn't modify nr_sectors or current_nr_sectors,
344 and hence can't be used for advancing request state settings on the
345 completion of partial transfers. The driver has to modify these fields
346 directly by hand.
347 This is because end_that_request_first only iterates over the bio list,
348 and always returns 0 if there are none associated with the request.
349 _last works OK in this case, and is not a problem, as I mentioned earlier
3521.3.1 Pre-built Commands
354A request can be created with a pre-built custom command to be sent directly
355to the device. The cmd block in the request structure has room for filling
356in the command bytes. (i.e rq->cmd is now 16 bytes in size, and meant for
357command pre-building, and the type of the request is now indicated
358through rq->flags instead of via rq->cmd)
360The request structure flags can be set up to indicate the type of request
361in such cases (REQ_PC: direct packet command passed to driver, REQ_BLOCK_PC:
362packet command issued via blk_do_rq, REQ_SPECIAL: special request).
364It can help to pre-build device commands for requests in advance.
365Drivers can now specify a request prepare function (q->prep_rq_fn) that the
366block layer would invoke to pre-build device commands for a given request,
367or perform other preparatory processing for the request. This is routine is
368called by elv_next_request(), i.e. typically just before servicing a request.
369(The prepare function would not be called for requests that have REQ_DONTPREP
373  Pre-building could possibly even be done early, i.e before placing the
374  request on the queue, rather than construct the command on the fly in the
375  driver while servicing the request queue when it may affect latencies in
376  interrupt context or responsiveness in general. One way to add early
377  pre-building would be to do it whenever we fail to merge on a request.
378  Now REQ_NOMERGE is set in the request flags to skip this one in the future,
379  which means that it will not change before we feed it to the device. So
380  the pre-builder hook can be invoked there.
3832. Flexible and generic but minimalist i/o structure/descriptor.
3852.1 Reason for a new structure and requirements addressed
387Prior to 2.5, buffer heads were used as the unit of i/o at the generic block
388layer, and the low level request structure was associated with a chain of
389buffer heads for a contiguous i/o request. This led to certain inefficiencies
390when it came to large i/o requests and readv/writev style operations, as it
391forced such requests to be broken up into small chunks before being passed
392on to the generic block layer, only to be merged by the i/o scheduler
393when the underlying device was capable of handling the i/o in one shot.
394Also, using the buffer head as an i/o structure for i/os that didn't originate
395from the buffer cache unnecessarily added to the weight of the descriptors
396which were generated for each such chunk.
398The following were some of the goals and expectations considered in the
399redesign of the block i/o data structure in 2.5.
401i. Should be appropriate as a descriptor for both raw and buffered i/o -
402    avoid cache related fields which are irrelevant in the direct/page i/o path,
403    or filesystem block size alignment restrictions which may not be relevant
404    for raw i/o.
405ii. Ability to represent high-memory buffers (which do not have a virtual
406    address mapping in kernel address space).
407iii.Ability to represent large i/os w/o unnecessarily breaking them up (i.e
408    greater than PAGE_SIZE chunks in one shot)
409iv. At the same time, ability to retain independent identity of i/os from
410    different sources or i/o units requiring individual completion (e.g. for
411    latency reasons)
412v. Ability to represent an i/o involving multiple physical memory segments
413    (including non-page aligned page fragments, as specified via readv/writev)
414    without unnecessarily breaking it up, if the underlying device is capable of
415    handling it.
416vi. Preferably should be based on a memory descriptor structure that can be
417    passed around different types of subsystems or layers, maybe even
418    networking, without duplication or extra copies of data/descriptor fields
419    themselves in the process
420vii.Ability to handle the possibility of splits/merges as the structure passes
421    through layered drivers (lvm, md, evms), with minimal overhead.
423The solution was to define a new structure (bio) for the block layer,
424instead of using the buffer head structure (bh) directly, the idea being
425avoidance of some associated baggage and limitations. The bio structure
426is uniformly used for all i/o at the block layer ; it forms a part of the
427bh structure for buffered i/o, and in the case of raw/direct i/o kiobufs are
428mapped to bio structures.
4302.2 The bio struct
432The bio structure uses a vector representation pointing to an array of tuples
433of <page, offset, len> to describe the i/o buffer, and has various other
434fields describing i/o parameters and state that needs to be maintained for
435performing the i/o.
437Notice that this representation means that a bio has no virtual address
438mapping at all (unlike buffer heads).
440struct bio_vec {
441       struct page *bv_page;
442       unsigned short bv_len;
443       unsigned short bv_offset;
447 * main unit of I/O for the block layer and lower layers (ie drivers)
448 */
449struct bio {
450       sector_t bi_sector;
451       struct bio *bi_next; /* request queue link */
452       struct block_device *bi_bdev; /* target device */
453       unsigned long bi_flags; /* status, command, etc */
454       unsigned long bi_rw; /* low bits: r/w, high: priority */
456       unsigned int bi_vcnt; /* how may bio_vec's */
457       unsigned int bi_idx; /* current index into bio_vec array */
459       unsigned int bi_size; /* total size in bytes */
460       unsigned short bi_phys_segments; /* segments after physaddr coalesce*/
461       unsigned short bi_hw_segments; /* segments after DMA remapping */
462       unsigned int bi_max; /* max bio_vecs we can hold
463                                        used as index into pool */
464       struct bio_vec *bi_io_vec; /* the actual vec list */
465       bio_end_io_t *bi_end_io; /* bi_end_io (bio) */
466       atomic_t bi_cnt; /* pin count: free when it hits zero */
467       void *bi_private;
468       bio_destructor_t *bi_destructor; /* bi_destructor (bio) */
471With this multipage bio design:
473- Large i/os can be sent down in one go using a bio_vec list consisting
474  of an array of <page, offset, len> fragments (similar to the way fragments
475  are represented in the zero-copy network code)
476- Splitting of an i/o request across multiple devices (as in the case of
477  lvm or raid) is achieved by cloning the bio (where the clone points to
478  the same bi_io_vec array, but with the index and size accordingly modified)
479- A linked list of bios is used as before for unrelated merges (*) - this
480  avoids reallocs and makes independent completions easier to handle.
481- Code that traverses the req list can find all the segments of a bio
482  by using rq_for_each_segment. This handles the fact that a request
483  has multiple bios, each of which can have multiple segments.
484- Drivers which can't process a large bio in one shot can use the bi_idx
485  field to keep track of the next bio_vec entry to process.
486  (e.g a 1MB bio_vec needs to be handled in max 128kB chunks for IDE)
487  [TBD: Should preferably also have a bi_voffset and bi_vlen to avoid modifying
488   bi_offset an len fields]
490(*) unrelated merges -- a request ends up containing two or more bios that
491    didn't originate from the same place.
493bi_end_io() i/o callback gets called on i/o completion of the entire bio.
495At a lower level, drivers build a scatter gather list from the merged bios.
496The scatter gather list is in the form of an array of <page, offset, len>
497entries with their corresponding dma address mappings filled in at the
498appropriate time. As an optimization, contiguous physical pages can be
499covered by a single entry where <page> refers to the first page and <len>
500covers the range of pages (up to 16 contiguous pages could be covered this
501way). There is a helper routine (blk_rq_map_sg) which drivers can use to build
502the sg list.
504Note: Right now the only user of bios with more than one page is ll_rw_kio,
505which in turn means that only raw I/O uses it (direct i/o may not work
506right now). The intent however is to enable clustering of pages etc to
507become possible. The pagebuf abstraction layer from SGI also uses multi-page
508bios, but that is currently not included in the stock development kernels.
509The same is true of Andrew Morton's work-in-progress multipage bio writeout
510and readahead patches.
5122.3 Changes in the Request Structure
514The request structure is the structure that gets passed down to low level
515drivers. The block layer make_request function builds up a request structure,
516places it on the queue and invokes the drivers request_fn. The driver makes
517use of block layer helper routine elv_next_request to pull the next request
518off the queue. Control or diagnostic functions might bypass block and directly
519invoke underlying driver entry points passing in a specially constructed
520request structure.
522Only some relevant fields (mainly those which changed or may be referred
523to in some of the discussion here) are listed below, not necessarily in
524the order in which they occur in the structure (see include/linux/blkdev.h)
525Refer to Documentation/block/request.txt for details about all the request
526structure fields and a quick reference about the layers which are
527supposed to use or modify those fields.
529struct request {
530    struct list_head queuelist; /* Not meant to be directly accessed by
531                    the driver.
532                    Used by q->elv_next_request_fn
533                    rq->queue is gone
534                    */
535    .
536    .
537    unsigned char cmd[16]; /* prebuilt command data block */
538    unsigned long flags; /* also includes earlier rq->cmd settings */
539    .
540    .
541    sector_t sector; /* this field is now of type sector_t instead of int
542                preparation for 64 bit sectors */
543    .
544    .
546    /* Number of scatter-gather DMA addr+len pairs after
547     * physical address coalescing is performed.
548     */
549    unsigned short nr_phys_segments;
551    /* Number of scatter-gather addr+len pairs after
552     * physical and DMA remapping hardware coalescing is performed.
553     * This is the number of scatter-gather entries the driver
554     * will actually have to deal with after DMA mapping is done.
555     */
556    unsigned short nr_hw_segments;
558    /* Various sector counts */
559    unsigned long nr_sectors; /* no. of sectors left: driver modifiable */
560    unsigned long hard_nr_sectors; /* block internal copy of above */
561    unsigned int current_nr_sectors; /* no. of sectors left in the
562                       current segment:driver modifiable */
563    unsigned long hard_cur_sectors; /* block internal copy of the above */
564    .
565    .
566    int tag; /* command tag associated with request */
567    void *special; /* same as before */
568    char *buffer; /* valid only for low memory buffers up to
569             current_nr_sectors */
570    .
571    .
572    struct bio *bio, *biotail; /* bio list instead of bh */
573    struct request_list *rl;
576See the rq_flag_bits definitions for an explanation of the various flags
577available. Some bits are used by the block layer or i/o scheduler.
579The behaviour of the various sector counts are almost the same as before,
580except that since we have multi-segment bios, current_nr_sectors refers
581to the numbers of sectors in the current segment being processed which could
582be one of the many segments in the current bio (i.e i/o completion unit).
583The nr_sectors value refers to the total number of sectors in the whole
584request that remain to be transferred (no change). The purpose of the
585hard_xxx values is for block to remember these counts every time it hands
586over the request to the driver. These values are updated by block on
587end_that_request_first, i.e. every time the driver completes a part of the
588transfer and invokes block end*request helpers to mark this. The
589driver should not modify these values. The block layer sets up the
590nr_sectors and current_nr_sectors fields (based on the corresponding
591hard_xxx values and the number of bytes transferred) and updates it on
592every transfer that invokes end_that_request_first. It does the same for the
593buffer, bio, bio->bi_idx fields too.
595The buffer field is just a virtual address mapping of the current segment
596of the i/o buffer in cases where the buffer resides in low-memory. For high
597memory i/o, this field is not valid and must not be used by drivers.
599Code that sets up its own request structures and passes them down to
600a driver needs to be careful about interoperation with the block layer helper
601functions which the driver uses. (Section 1.3)
6033. Using bios
6053.1 Setup/Teardown
607There are routines for managing the allocation, and reference counting, and
608freeing of bios (bio_alloc, bio_get, bio_put).
610This makes use of Ingo Molnar's mempool implementation, which enables
611subsystems like bio to maintain their own reserve memory pools for guaranteed
612deadlock-free allocations during extreme VM load. For example, the VM
613subsystem makes use of the block layer to writeout dirty pages in order to be
614able to free up memory space, a case which needs careful handling. The
615allocation logic draws from the preallocated emergency reserve in situations
616where it cannot allocate through normal means. If the pool is empty and it
617can wait, then it would trigger action that would help free up memory or
618replenish the pool (without deadlocking) and wait for availability in the pool.
619If it is in IRQ context, and hence not in a position to do this, allocation
620could fail if the pool is empty. In general mempool always first tries to
621perform allocation without having to wait, even if it means digging into the
622pool as long it is not less that 50% full.
624On a free, memory is released to the pool or directly freed depending on
625the current availability in the pool. The mempool interface lets the
626subsystem specify the routines to be used for normal alloc and free. In the
627case of bio, these routines make use of the standard slab allocator.
629The caller of bio_alloc is expected to taken certain steps to avoid
630deadlocks, e.g. avoid trying to allocate more memory from the pool while
631already holding memory obtained from the pool.
632[TBD: This is a potential issue, though a rare possibility
633 in the bounce bio allocation that happens in the current code, since
634 it ends up allocating a second bio from the same pool while
635 holding the original bio ]
637Memory allocated from the pool should be released back within a limited
638amount of time (in the case of bio, that would be after the i/o is completed).
639This ensures that if part of the pool has been used up, some work (in this
640case i/o) must already be in progress and memory would be available when it
641is over. If allocating from multiple pools in the same code path, the order
642or hierarchy of allocation needs to be consistent, just the way one deals
643with multiple locks.
645The bio_alloc routine also needs to allocate the bio_vec_list (bvec_alloc())
646for a non-clone bio. There are the 6 pools setup for different size biovecs,
647so bio_alloc(gfp_mask, nr_iovecs) will allocate a vec_list of the
648given size from these slabs.
650The bi_destructor() routine takes into account the possibility of the bio
651having originated from a different source (see later discussions on
652n/w to block transfers and kvec_cb)
654The bio_get() routine may be used to hold an extra reference on a bio prior
655to i/o submission, if the bio fields are likely to be accessed after the
656i/o is issued (since the bio may otherwise get freed in case i/o completion
657happens in the meantime).
659The bio_clone() routine may be used to duplicate a bio, where the clone
660shares the bio_vec_list with the original bio (i.e. both point to the
661same bio_vec_list). This would typically be used for splitting i/o requests
662in lvm or md.
6643.2 Generic bio helper Routines
6663.2.1 Traversing segments and completion units in a request
668The macro rq_for_each_segment() should be used for traversing the bios
669in the request list (drivers should avoid directly trying to do it
670themselves). Using these helpers should also make it easier to cope
671with block changes in the future.
673    struct req_iterator iter;
674    rq_for_each_segment(bio_vec, rq, iter)
675        /* bio_vec is now current segment */
677I/O completion callbacks are per-bio rather than per-segment, so drivers
678that traverse bio chains on completion need to keep that in mind. Drivers
679which don't make a distinction between segments and completion units would
680need to be reorganized to support multi-segment bios.
6823.2.2 Setting up DMA scatterlists
684The blk_rq_map_sg() helper routine would be used for setting up scatter
685gather lists from a request, so a driver need not do it on its own.
687    nr_segments = blk_rq_map_sg(q, rq, scatterlist);
689The helper routine provides a level of abstraction which makes it easier
690to modify the internals of request to scatterlist conversion down the line
691without breaking drivers. The blk_rq_map_sg routine takes care of several
692things like collapsing physically contiguous segments (if QUEUE_FLAG_CLUSTER
693is set) and correct segment accounting to avoid exceeding the limits which
694the i/o hardware can handle, based on various queue properties.
696- Prevents a clustered segment from crossing a 4GB mem boundary
697- Avoids building segments that would exceed the number of physical
698  memory segments that the driver can handle (phys_segments) and the
699  number that the underlying hardware can handle at once, accounting for
700  DMA remapping (hw_segments) (i.e. IOMMU aware limits).
702Routines which the low level driver can use to set up the segment limits:
704blk_queue_max_hw_segments() : Sets an upper limit of the maximum number of
705hw data segments in a request (i.e. the maximum number of address/length
706pairs the host adapter can actually hand to the device at once)
708blk_queue_max_phys_segments() : Sets an upper limit on the maximum number
709of physical data segments in a request (i.e. the largest sized scatter list
710a driver could handle)
7123.2.3 I/O completion
714The existing generic block layer helper routines end_request,
715end_that_request_first and end_that_request_last can be used for i/o
716completion (and setting things up so the rest of the i/o or the next
717request can be kicked of) as before. With the introduction of multi-page
718bio support, end_that_request_first requires an additional argument indicating
719the number of sectors completed.
7213.2.4 Implications for drivers that do not interpret bios (don't handle
722 multiple segments)
724Drivers that do not interpret bios e.g those which do not handle multiple
725segments and do not support i/o into high memory addresses (require bounce
726buffers) and expect only virtually mapped buffers, can access the rq->buffer
727field. As before the driver should use current_nr_sectors to determine the
728size of remaining data in the current segment (that is the maximum it can
729transfer in one go unless it interprets segments), and rely on the block layer
730end_request, or end_that_request_first/last to take care of all accounting
731and transparent mapping of the next bio segment when a segment boundary
732is crossed on completion of a transfer. (The end*request* functions should
733be used if only if the request has come down from block/bio path, not for
734direct access requests which only specify rq->buffer without a valid rq->bio)
7363.2.5 Generic request command tagging
7383.2.5.1 Tag helpers
740Block now offers some simple generic functionality to help support command
741queueing (typically known as tagged command queueing), ie manage more than
742one outstanding command on a queue at any given time.
744    blk_queue_init_tags(struct request_queue *q, int depth)
746    Initialize internal command tagging structures for a maximum
747    depth of 'depth'.
749    blk_queue_free_tags((struct request_queue *q)
751    Teardown tag info associated with the queue. This will be done
752    automatically by block if blk_queue_cleanup() is called on a queue
753    that is using tagging.
755The above are initialization and exit management, the main helpers during
756normal operations are:
758    blk_queue_start_tag(struct request_queue *q, struct request *rq)
760    Start tagged operation for this request. A free tag number between
761    0 and 'depth' is assigned to the request (rq->tag holds this number),
762    and 'rq' is added to the internal tag management. If the maximum depth
763    for this queue is already achieved (or if the tag wasn't started for
764    some other reason), 1 is returned. Otherwise 0 is returned.
766    blk_queue_end_tag(struct request_queue *q, struct request *rq)
768    End tagged operation on this request. 'rq' is removed from the internal
769    book keeping structures.
771To minimize struct request and queue overhead, the tag helpers utilize some
772of the same request members that are used for normal request queue management.
773This means that a request cannot both be an active tag and be on the queue
774list at the same time. blk_queue_start_tag() will remove the request, but
775the driver must remember to call blk_queue_end_tag() before signalling
776completion of the request to the block layer. This means ending tag
777operations before calling end_that_request_last()! For an example of a user
778of these helpers, see the IDE tagged command queueing support.
780Certain hardware conditions may dictate a need to invalidate the block tag
781queue. For instance, on IDE any tagged request error needs to clear both
782the hardware and software block queue and enable the driver to sanely restart
783all the outstanding requests. There's a third helper to do that:
785    blk_queue_invalidate_tags(struct request_queue *q)
787    Clear the internal block tag queue and re-add all the pending requests
788    to the request queue. The driver will receive them again on the
789    next request_fn run, just like it did the first time it encountered
790    them.
7923.2.5.2 Tag info
794Some block functions exist to query current tag status or to go from a
795tag number to the associated request. These are, in no particular order:
797    blk_queue_tagged(q)
799    Returns 1 if the queue 'q' is using tagging, 0 if not.
801    blk_queue_tag_request(q, tag)
803    Returns a pointer to the request associated with tag 'tag'.
805    blk_queue_tag_depth(q)
807    Return current queue depth.
809    blk_queue_tag_queue(q)
811    Returns 1 if the queue can accept a new queued command, 0 if we are
812    at the maximum depth already.
814    blk_queue_rq_tagged(rq)
816    Returns 1 if the request 'rq' is tagged.
8183.2.5.2 Internal structure
820Internally, block manages tags in the blk_queue_tag structure:
822    struct blk_queue_tag {
823        struct request **tag_index; /* array or pointers to rq */
824        unsigned long *tag_map; /* bitmap of free tags */
825        struct list_head busy_list; /* fifo list of busy tags */
826        int busy; /* queue depth */
827        int max_depth; /* max queue depth */
828    };
830Most of the above is simple and straight forward, however busy_list may need
831a bit of explaining. Normally we don't care too much about request ordering,
832but in the event of any barrier requests in the tag queue we need to ensure
833that requests are restarted in the order they were queue. This may happen
834if the driver needs to use blk_queue_invalidate_tags().
836Tagging also defines a new request flag, REQ_QUEUED. This is set whenever
837a request is currently tagged. You should not use this flag directly,
838blk_rq_tagged(rq) is the portable way to do so.
8403.3 I/O Submission
842The routine submit_bio() is used to submit a single io. Higher level i/o
843routines make use of this:
845(a) Buffered i/o:
846The routine submit_bh() invokes submit_bio() on a bio corresponding to the
847bh, allocating the bio if required. ll_rw_block() uses submit_bh() as before.
849(b) Kiobuf i/o (for raw/direct i/o):
850The ll_rw_kio() routine breaks up the kiobuf into page sized chunks and
851maps the array to one or more multi-page bios, issuing submit_bio() to
852perform the i/o on each of these.
854The embedded bh array in the kiobuf structure has been removed and no
855preallocation of bios is done for kiobufs. [The intent is to remove the
856blocks array as well, but it's currently in there to kludge around direct i/o.]
857Thus kiobuf allocation has switched back to using kmalloc rather than vmalloc.
861 A single kiobuf structure is assumed to correspond to a contiguous range
862 of data, so brw_kiovec() invokes ll_rw_kio for each kiobuf in a kiovec.
863 So right now it wouldn't work for direct i/o on non-contiguous blocks.
864 This is to be resolved. The eventual direction is to replace kiobuf
865 by kvec's.
867 Badari Pulavarty has a patch to implement direct i/o correctly using
868 bio and kvec.
871(c) Page i/o:
872Todo/Under discussion:
874 Andrew Morton's multi-page bio patches attempt to issue multi-page
875 writeouts (and reads) from the page cache, by directly building up
876 large bios for submission completely bypassing the usage of buffer
877 heads. This work is still in progress.
879 Christoph Hellwig had some code that uses bios for page-io (rather than
880 bh). This isn't included in bio as yet. Christoph was also working on a
881 design for representing virtual/real extents as an entity and modifying
882 some of the address space ops interfaces to utilize this abstraction rather
883 than buffer_heads. (This is somewhat along the lines of the SGI XFS pagebuf
884 abstraction, but intended to be as lightweight as possible).
886(d) Direct access i/o:
887Direct access requests that do not contain bios would be submitted differently
888as discussed earlier in section 1.3.
892  Kvec i/o:
894  Ben LaHaise's aio code uses a slightly different structure instead
895  of kiobufs, called a kvec_cb. This contains an array of <page, offset, len>
896  tuples (very much like the networking code), together with a callback function
897  and data pointer. This is embedded into a brw_cb structure when passed
898  to brw_kvec_async().
900  Now it should be possible to directly map these kvecs to a bio. Just as while
901  cloning, in this case rather than PRE_BUILT bio_vecs, we set the bi_io_vec
902  array pointer to point to the veclet array in kvecs.
904  TBD: In order for this to work, some changes are needed in the way multi-page
905  bios are handled today. The values of the tuples in such a vector passed in
906  from higher level code should not be modified by the block layer in the course
907  of its request processing, since that would make it hard for the higher layer
908  to continue to use the vector descriptor (kvec) after i/o completes. Instead,
909  all such transient state should either be maintained in the request structure,
910  and passed on in some way to the endio completion routine.
9134. The I/O scheduler
914I/O scheduler, a.k.a. elevator, is implemented in two layers. Generic dispatch
915queue and specific I/O schedulers. Unless stated otherwise, elevator is used
916to refer to both parts and I/O scheduler to specific I/O schedulers.
918Block layer implements generic dispatch queue in block/*.c.
919The generic dispatch queue is responsible for properly ordering barrier
920requests, requeueing, handling non-fs requests and all other subtleties.
922Specific I/O schedulers are responsible for ordering normal filesystem
923requests. They can also choose to delay certain requests to improve
924throughput or whatever purpose. As the plural form indicates, there are
925multiple I/O schedulers. They can be built as modules but at least one should
926be built inside the kernel. Each queue can choose different one and can also
927change to another one dynamically.
929A block layer call to the i/o scheduler follows the convention elv_xxx(). This
930calls elevator_xxx_fn in the elevator switch (block/elevator.c). Oh, xxx
931and xxx might not match exactly, but use your imagination. If an elevator
932doesn't implement a function, the switch does nothing or some minimal house
933keeping work.
9354.1. I/O scheduler API
937The functions an elevator may implement are: (* are mandatory)
938elevator_merge_fn called to query requests for merge with a bio
940elevator_merge_req_fn called when two requests get merged. the one
941                which gets merged into the other one will be
942                never seen by I/O scheduler again. IOW, after
943                being merged, the request is gone.
945elevator_merged_fn called when a request in the scheduler has been
946                involved in a merge. It is used in the deadline
947                scheduler for example, to reposition the request
948                if its sorting order has changed.
950elevator_allow_merge_fn called whenever the block layer determines
951                that a bio can be merged into an existing
952                request safely. The io scheduler may still
953                want to stop a merge at this point if it
954                results in some sort of conflict internally,
955                this hook allows it to do that.
957elevator_dispatch_fn* fills the dispatch queue with ready requests.
958                I/O schedulers are free to postpone requests by
959                not filling the dispatch queue unless @force
960                is non-zero. Once dispatched, I/O schedulers
961                are not allowed to manipulate the requests -
962                they belong to generic dispatch queue.
964elevator_add_req_fn* called to add a new request into the scheduler
967elevator_latter_req_fn These return the request before or after the
968                one specified in disk sort order. Used by the
969                block layer to find merge possibilities.
971elevator_completed_req_fn called when a request is completed.
973elevator_may_queue_fn returns true if the scheduler wants to allow the
974                current context to queue a new request even if
975                it is over the queue limit. This must be used
976                very carefully!!
979elevator_put_req_fn Must be used to allocate and free any elevator
980                specific storage for a request.
982elevator_activate_req_fn Called when device driver first sees a request.
983                I/O schedulers can use this callback to
984                determine when actual execution of a request
985                starts.
986elevator_deactivate_req_fn Called when device driver decides to delay
987                a request by requeueing it.
990elevator_exit_fn Allocate and free any elevator specific storage
991                for a queue.
9934.2 Request flows seen by I/O schedulers
994All requests seen by I/O schedulers strictly follow one of the following three
997 set_req_fn ->
999 i. add_req_fn -> (merged_fn ->)* -> dispatch_fn -> activate_req_fn ->
1000      (deactivate_req_fn -> activate_req_fn ->)* -> completed_req_fn
1001 ii. add_req_fn -> (merged_fn ->)* -> merge_req_fn
1002 iii. [none]
1004 -> put_req_fn
10064.3 I/O scheduler implementation
1007The generic i/o scheduler algorithm attempts to sort/merge/batch requests for
1008optimal disk scan and request servicing performance (based on generic
1009principles and device capabilities), optimized for:
1010i. improved throughput
1011ii. improved latency
1012iii. better utilization of h/w & CPU time
1016i. Binary tree
1017AS and deadline i/o schedulers use red black binary trees for disk position
1018sorting and searching, and a fifo linked list for time-based searching. This
1019gives good scalability and good availability of information. Requests are
1020almost always dispatched in disk sort order, so a cache is kept of the next
1021request in sort order to prevent binary tree lookups.
1023This arrangement is not a generic block layer characteristic however, so
1024elevators may implement queues as they please.
1026ii. Merge hash
1027AS and deadline use a hash table indexed by the last sector of a request. This
1028enables merging code to quickly look up "back merge" candidates, even when
1029multiple I/O streams are being performed at once on one disk.
1031"Front merges", a new request being merged at the front of an existing request,
1032are far less common than "back merges" due to the nature of most I/O patterns.
1033Front merges are handled by the binary trees in AS and deadline schedulers.
1035iii. Plugging the queue to batch requests in anticipation of opportunities for
1036     merge/sort optimizations
1038Plugging is an approach that the current i/o scheduling algorithm resorts to so
1039that it collects up enough requests in the queue to be able to take
1040advantage of the sorting/merging logic in the elevator. If the
1041queue is empty when a request comes in, then it plugs the request queue
1042(sort of like plugging the bath tub of a vessel to get fluid to build up)
1043till it fills up with a few more requests, before starting to service
1044the requests. This provides an opportunity to merge/sort the requests before
1045passing them down to the device. There are various conditions when the queue is
1046unplugged (to open up the flow again), either through a scheduled task or
1047could be on demand. For example wait_on_buffer sets the unplugging going
1048through sync_buffer() running blk_run_address_space(mapping). Or the caller
1049can do it explicity through blk_unplug(bdev). So in the read case,
1050the queue gets explicitly unplugged as part of waiting for completion on that
1051buffer. For page driven IO, the address space ->sync_page() takes care of
1052doing the blk_run_address_space().
1055  This is kind of controversial territory, as it's not clear if plugging is
1056  always the right thing to do. Devices typically have their own queues,
1057  and allowing a big queue to build up in software, while letting the device be
1058  idle for a while may not always make sense. The trick is to handle the fine
1059  balance between when to plug and when to open up. Also now that we have
1060  multi-page bios being queued in one shot, we may not need to wait to merge
1061  a big request from the broken up pieces coming by.
10634.4 I/O contexts
1064I/O contexts provide a dynamically allocated per process data area. They may
1065be used in I/O schedulers, and in the block layer (could be used for IO statis,
1066priorities for example). See *io_context in block/ll_rw_blk.c, and as-iosched.c
1067for an example of usage in an i/o scheduler.
10705. Scalability related changes
10725.1 Granular Locking: io_request_lock replaced by a per-queue lock
1074The global io_request_lock has been removed as of 2.5, to avoid
1075the scalability bottleneck it was causing, and has been replaced by more
1076granular locking. The request queue structure has a pointer to the
1077lock to be used for that queue. As a result, locking can now be
1078per-queue, with a provision for sharing a lock across queues if
1079necessary (e.g the scsi layer sets the queue lock pointers to the
1080corresponding adapter lock, which results in a per host locking
1081granularity). The locking semantics are the same, i.e. locking is
1082still imposed by the block layer, grabbing the lock before
1083request_fn execution which it means that lots of older drivers
1084should still be SMP safe. Drivers are free to drop the queue
1085lock themselves, if required. Drivers that explicitly used the
1086io_request_lock for serialization need to be modified accordingly.
1087Usually it's as easy as adding a global lock:
1089    static DEFINE_SPINLOCK(my_driver_lock);
1091and passing the address to that lock to blk_init_queue().
10935.2 64 bit sector numbers (sector_t prepares for 64 bit support)
1095The sector number used in the bio structure has been changed to sector_t,
1096which could be defined as 64 bit in preparation for 64 bit sector support.
10986. Other Changes/Implications
11006.1 Partition re-mapping handled by the generic block layer
1102In 2.5 some of the gendisk/partition related code has been reorganized.
1103Now the generic block layer performs partition-remapping early and thus
1104provides drivers with a sector number relative to whole device, rather than
1105having to take partition number into account in order to arrive at the true
1106sector number. The routine blk_partition_remap() is invoked by
1107generic_make_request even before invoking the queue specific make_request_fn,
1108so the i/o scheduler also gets to operate on whole disk sector numbers. This
1109should typically not require changes to block drivers, it just never gets
1110to invoke its own partition sector offset calculations since all bios
1111sent are offset from the beginning of the device.
11147. A Few Tips on Migration of older drivers
1116Old-style drivers that just use CURRENT and ignores clustered requests,
1117may not need much change. The generic layer will automatically handle
1118clustered requests, multi-page bios, etc for the driver.
1120For a low performance driver or hardware that is PIO driven or just doesn't
1121support scatter-gather changes should be minimal too.
1123The following are some points to keep in mind when converting old drivers
1124to bio.
1126Drivers should use elv_next_request to pick up requests and are no longer
1127supposed to handle looping directly over the request list.
1128(struct request->queue has been removed)
1130Now end_that_request_first takes an additional number_of_sectors argument.
1131It used to handle always just the first buffer_head in a request, now
1132it will loop and handle as many sectors (on a bio-segment granularity)
1133as specified.
1135Now bh->b_end_io is replaced by bio->bi_end_io, but most of the time the
1136right thing to use is bio_endio(bio, uptodate) instead.
1138If the driver is dropping the io_request_lock from its request_fn strategy,
1139then it just needs to replace that with q->queue_lock instead.
1141As described in Sec 1.1, drivers can set max sector size, max segment size
1142etc per queue now. Drivers that used to define their own merge functions i
1143to handle things like this can now just use the blk_queue_* functions at
1144blk_init_queue time.
1146Drivers no longer have to map a {partition, sector offset} into the
1147correct absolute location anymore, this is done by the block layer, so
1148where a driver received a request ala this before:
1150    rq->rq_dev = mk_kdev(3, 5); /* /dev/hda5 */
1151    rq->sector = 0; /* first sector on hda5 */
1153  it will now see
1155    rq->rq_dev = mk_kdev(3, 0); /* /dev/hda */
1156    rq->sector = 123128; /* offset from start of disk */
1158As mentioned, there is no virtual mapping of a bio. For DMA, this is
1159not a problem as the driver probably never will need a virtual mapping.
1160Instead it needs a bus mapping (dma_map_page for a single segment or
1161use dma_map_sg for scatter gather) to be able to ship it to the driver. For
1162PIO drivers (or drivers that need to revert to PIO transfer once in a
1163while (IDE for example)), where the CPU is doing the actual data
1164transfer a virtual mapping is needed. If the driver supports highmem I/O,
1165(Sec 1.1, (ii) ) it needs to use __bio_kmap_atomic and bio_kmap_irq to
1166temporarily map a bio into the virtual address space.
11698. Prior/Related/Impacted patches
11718.1. Earlier kiobuf patches (sct/axboe/chait/hch/mkp)
1172- orig kiobuf & raw i/o patches (now in 2.4 tree)
1173- direct kiobuf based i/o to devices (no intermediate bh's)
1174- page i/o using kiobuf
1175- kiobuf splitting for lvm (mkp)
1176- elevator support for kiobuf request merging (axboe)
11778.2. Zero-copy networking (Dave Miller)
11788.3. SGI XFS - pagebuf patches - use of kiobufs
11798.4. Multi-page pioent patch for bio (Christoph Hellwig)
11808.5. Direct i/o implementation (Andrea Arcangeli) since 2.4.10-pre11
11818.6. Async i/o implementation patch (Ben LaHaise)
11828.7. EVMS layering design (IBM EVMS team)
11838.8. Larger page cache size patch (Ben LaHaise) and
1184     Large page size (Daniel Phillips)
1185    => larger contiguous physical memory buffers
11868.9. VM reservations patch (Ben LaHaise)
11878.10. Write clustering patches ? (Marcelo/Quintela/Riel ?)
11888.11. Block device in page cache patch (Andrea Archangeli) - now in 2.4.10+
11898.12. Multiple block-size transfers for faster raw i/o (Shailabh Nagar,
1190      Badari)
11918.13 Priority based i/o scheduler - prepatches (Arjan van de Ven)
11928.14 IDE Taskfile i/o patch (Andre Hedrick)
11938.15 Multi-page writeout and readahead patches (Andrew Morton)
11948.16 Direct i/o patches for 2.5 using kvec and bio (Badari Pulavarthy)
11969. Other References:
11989.1 The Splice I/O Model - Larry McVoy (and subsequent discussions on lkml,
1199and Linus' comments - Jan 2001)
12009.2 Discussions about kiobuf and bh design on lkml between sct, linus, alan
1201et al - Feb-March 2001 (many of the initial thoughts that led to bio were
1202brought up in this discussion thread)
12039.3 Discussions on mempool on lkml - Dec 2001.

Archive Download this file